TechSurge: Deep Tech Podcast

Cisco's Jeetu Patel on AI Security, Token Economics, and Infrastructure Planning Cycles

1h 1m Listen to the original Episode page
AI SecurityToken EconomicsCiscoInfrastructureSiliconIdentity ManagementData CentersEnterprise AI

Summary

Cisco President Jeetu Patel argues that token exhaustion is a critical security vulnerability, framing token efficiency as a discipline necessary to prevent adversaries from disabling AI agents. He details Cisco's strategic pivot from a holding company to an integrated platform spanning silicon, photonics, and software, citing acquisitions like Acacia and the development of merchant silicon as key to serving hyperscalers. Patel highlights that agent adoption lags behind chatbots due to trust issues regarding sensitive data access, necessitating robust identity management and observability layers. He also contrasts the AI infrastructure buildout with the 5G era, noting that power availability and cost of capital are now the primary constraints rather than demand lag. Finally, Patel outlines Cisco's diverse planning horizons across the tech stack, from five-year silicon cycles to one-week agent cycles, as a core competitive moat.

Key takeaways

  1. 1
    your security posture is directly proportionate to your ability to use agents to transform your business.

    Unidentified speaker (speaker_0) · 21:46

Episode sections

  1. 0:00 – 10:32

    Token Security as a Prerequisite for AI Adoption

    Speakers introduce the concept that running out of tokens is akin to running out of ammunition in war, establishing security as a prerequisite for relevance in AI. Cisco President Jeetu Patel notes that trust, not just compute, is the gating factor for agent adoption, which is currently slower than chatbot adoption due to user hesitation to grant agents access to sensitive data like email and credit cards.

  2. 10:32 – 21:35

    Cisco's Strategic Pivot from Holding Company to Integrated Platform

    Patel explains Cisco's transformation from a 'holding company' with 250 acquisitions into a single integrated platform spanning silicon, photonics, and software. He cites the acquisition of Acacia for photonics and the strategic shift to building their own merchant silicon via Libra as key moves to serve hyperscalers, noting that silicon and optics now drive significant revenue from these customers.

  3. 21:35 – 25:11

    Tokenomics as a Security Discipline and Splunk's Role

    Patel argues that token efficiency has become a security discipline, where adversaries can exhaust an agent's token budget to leave it defenseless. He describes Cisco's stack, built on Splunk, as providing a universal registry for token usage, enabling algorithmic red teaming, agent behavior monitoring, and runtime enforcement to prevent lateral movement and token exhaustion attacks.

  4. 25:11 – 26:09

    Security and Token Economics in Agentic AI

    Speaker_0 argues that defending against machine-scale attacks requires an agent workforce, making token efficiency a critical security discipline. Without sufficient tokens, agents are defenseless, likened to running out of ammunition.

  5. 26:09 – 29:37

    Intelligent Routing and Model Diversity

    The speakers discuss the necessity of intelligent routing layers to manage multiple model providers. Speaker_0 predicts a perpetual need for a neutral third-party security layer that operates across all models and applications.

  6. 29:37 – 35:11

    Token Efficiency and Cost Reduction

    Speaker_0 highlights that older models (n-minus-5) are often more efficient and cheaper than newer ones. He notes that agents consumed seven times more tokens than humans by February, creating a demand for cost-effective routing.

  7. 35:11 – 44:58

    Measuring AI Output and Accountability

    The discussion shifts to measuring AI success via cost per task rather than just output. Speaker_0 emphasizes that humans must retain ultimate accountability for code, even if AI writes the majority of it, necessitating rigorous testing and simulation.

  8. 44:58 – 46:46

    Cisco's Trust Architecture and Identity

    Speaker_0 outlines a trust architecture based on visibility, observability, and identity. He mentions Cisco's acquisitions of Asterix and Whitefield to cover non-human and agent identities, ensuring every entity in the environment is monitored.

  9. 46:46 – 47:40

    Cisco's Networking Strategy: Scale Out vs Scale Up

    The conversation concludes with Cisco's focus on scale-out and scale-across networking architectures. Speaker_0 confirms that Cisco has not yet announced specific products for scale-up, focusing instead on connecting data centers across large distances.

  10. 47:40 – 50:05

    Scale-out silicon and enterprise data center expansion

    Speaker_0 describes Cisco's G300 scale-out chip, featuring ~250 trillion transistors on a 102 terabit, 100 sq mm package using 3nm process. He notes that while scale-up is becoming role-level, scale-out is currently focused on, with enterprises in financial services, oil & gas, pharma, and healthcare building their own AI data centers.

  11. 50:05 – 51:47

    Supply constraints: Power and cost of capital

    Speaker_0 argues that unlike the 5G buildout where demand lagged supply, AI supply is instantly consumed. The primary constraints are now available power (creating 'dark GPUs') and the cost of capital.

  12. 51:47 – 54:13

    Planning cycle variability across the stack

    Speaker_0 outlines Cisco's diverse planning horizons: Silicon/Photonics (5 years), Systems (24-36 months), OS (15-20 months), Platform (12-18 months), Models (3-4 months), Applications (months), and Agents (1 week). He identifies the ability to coordinate across these varying cycles as a key moat.

  13. 54:13 – 1:01:01

    Leadership culture and accountability

    Speaker_0 discusses his intrinsic motivation and leadership philosophy, emphasizing the importance of acknowledging mistakes, providing context to avoid misinterpretation, and surrounding oneself with people who provide honest feedback rather than just compliments.

Assets discussed

AssetKindStanceContext
NVIDIANVDAequityneutralIn fact, I I have talked about this in some of the other episodes where Nvidia

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

Google was processing more than 3.2 quadrillion tokens a month by May, seven times as many as a year earlier.

Unidentified speaker (speaker_1) ·0:05

How to verify: Check Google Cloud or Alphabet Inc. (GOOGL) quarterly earnings reports and investor presentations for Q2 2026 (released July 2026) for details on AI infrastructure scale and token processing volumes. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Last fiscal year, Cisco took in more than $9 billion of AI infrastructure orders from the largest cloud providers.

Unidentified speaker (speaker_1) ·0:45

How to verify: Check Cisco Systems (CSCO) Form 10-K for the fiscal year ending July 2026, filed with the SEC on EDGAR, for revenue breakdowns by customer segment and product line. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Cisco briefly became the world's most valuable company during the dot-com boom.

Unidentified speaker (speaker_1) ·0:45

How to verify: Review historical market capitalization data from Bloomberg or Reuters for the year 2026 to confirm Cisco's peak valuation relative to Microsoft and Intel. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco had stopped innovating for about a decade to a decade and a half prior to Jeetu Patel's tenure.

Unidentified speaker (UNKNOWN) ·10:32

How to verify: Review Cisco's product release history and patent filings from 2026-2026 compared to the late 1990s to assess innovation velocity; cross-reference with Patel's start date at Cisco. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Cisco had 250 acquisitions.

Unidentified speaker (speaker_0) ·11:08

How to verify: Review Cisco's investor relations page or SEC filings for a list of acquisitions; verify the count against historical M&A records. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Cisco acquired a company called Libra to build its own merchant silicon.

Unidentified speaker (UNKNOWN) ·14:41

How to verify: Check Cisco's press releases or SEC filings for the acquisition of Libra Semiconductor, which occurred in 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Cisco acquired a company called Acacia for photonics.

Unidentified speaker (UNKNOWN) ·15:13

How to verify: Check Cisco's press releases or SEC filings for the acquisition of Acacia Communications, which occurred in 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Coherent optics and optical transceivers account for forty percent of Cisco's revenue from hyperscalers.

Unidentified speaker (speaker_3) ·15:31

How to verify: Check Cisco's quarterly earnings transcripts and investor presentations for the most recent quarters prior to the podcast date to verify revenue mix by product and customer segment. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco acquired Splunk.

Unidentified speaker (speaker_0) ·20:19

How to verify: Check Cisco's Form 8-K or proxy statements regarding the completion of the Splunk acquisition, which closed in 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

In April 2000, Microsoft, Intel, and Cisco each had a market cap of five hundred billion dollars.

Unidentified speaker (UNKNOWN) ·9:14

How to verify: Review historical stock price data from Yahoo Finance or Bloomberg for MSFT, INTC, and CSCO for April 2026 to verify market capitalization levels. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Agents consumed seven times more tokens than humans by February.

Unidentified speaker (speaker_0) ·33:33

How to verify: Verify via Cisco's internal usage reports or public statements from Cisco leadership regarding AI token consumption trends in early 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco booked more than $9 billion in AI infrastructure orders from the largest cloud providers last fiscal year.

Unidentified speaker (UNKNOWN) ·25:11

How to verify: Check Cisco's Q4 FY2026 earnings report or 10-K filing for AI infrastructure revenue and order backlog details. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Cisco acquired Asterix and Whitefield.

Unidentified speaker (speaker_0) ·46:28

How to verify: Verify via Cisco's press releases or SEC filings (Schedule 14A or 8-K) detailing acquisitions of Asterix and Whitefield. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Primary source

Anthropic claims 90% of Claude was written by Claude.

Unidentified speaker (UNKNOWN) ·41:30

How to verify: Check Anthropic's official blog posts or technical reports regarding Claude's development process and code generation statistics. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco has not announced anything on scale-up networking.

Unidentified speaker (speaker_3) ·47:21

How to verify: Verify via Cisco's product announcement history or press releases regarding scale-up networking solutions as of November 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Token efficiency is a security discipline because adversaries can exhaust token budgets.

Unidentified speaker (speaker_0) ·25:48

How to verify: This is a strategic claim; verify via Cisco's published security whitepapers or blog posts on agentic AI security and token economics. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Agents are slower to adopt than chatbots due to trust issues.

Unidentified speaker (speaker_3) ·25:23

How to verify: Verify via Cisco's market research reports or public statements on AI adoption trends and trust barriers. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco's AI infrastructure orders are part of a buildout comparable to the dot-com boom.

Unidentified speaker (UNKNOWN) ·25:11

How to verify: Verify via Cisco's earnings call transcripts or CEO/President remarks comparing current AI infrastructure spending to historical periods. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Mechanistic interpretability is a field of study in its early days.

Unidentified speaker (speaker_0) ·27:17

How to verify: Verify via academic literature or industry reports on the state of mechanistic interpretability research as of 2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Intelligent routing layers will be invested in by everyone, whether bought or built.

Unidentified speaker (UNKNOWN) ·28:20

How to verify: This is a forward-looking assertion; verify via Cisco's strategic outlook documents or industry analyst reports on AI routing trends. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco's G300 chip has close to a quarter trillion transistors, is 102 terabit, on a 100 square millimeter package substrate, and uses a 3nm process.

Unidentified speaker (UNKNOWN) ·48:24

How to verify: Check Cisco's official product announcements or press releases for the G300 chip specifications; verify transistor count and process node via semiconductor industry reports (e.g., TechInsights). For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Cisco added debuffering technology to prevent restarting training runs across two data centers if a packet is dropped.

Unidentified speaker (UNKNOWN) ·49:10

How to verify: Review Cisco's technical documentation or white papers on their scale-out networking solutions for mentions of debuffering or packet loss mitigation in multi-datacenter training scenarios. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Any AI supply made available is instantly consumed, unlike the 5G buildout where demand lagged supply.

Unidentified speaker (speaker_0) ·50:18

How to verify: Compare cloud provider capital expenditure reports and utilization metrics from 2026-2026 against telecom infrastructure utilization data from 2026-2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Silicon and Photonics planning cycles are 5 years; Systems business is 24-36 months; OS is 15-20 months; Platform is 12-18 months; Model business is 3-4 months; Applications are a couple of months; Agents are 1 week.

Unidentified speaker (UNKNOWN) ·52:11

How to verify: Verify against Cisco's investor presentations or earnings call transcripts detailing their product development roadmaps and planning horizons. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Chip manufacturing requires a six-year planning cycle for fabs, conflicting with the one-week planning cycle of the agent market.

Unidentified speaker (speaker_0) ·51:19

How to verify: Review semiconductor industry analyses on fab construction lead times and planning requirements versus AI application development lifecycles. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify Cisco's hyperscaler revenue breakdown and the specific contribution of Acacia and coherent optics by reviewing Cisco's latest 10-K or 10-Q SEC filings.
  • Check Cisco's quarterly earnings reports for details on the 'single platform' strategy and the financial impact of the 250+ acquisitions mentioned.
  • Research the technical specifications of the Cisco G300 chip, including transistor count and process node, via Cisco's official product announcements or semiconductor industry reports.
  • Investigate the 'dark GPU' phenomenon by reviewing power grid capacity reports and data center energy consumption forecasts from energy regulators or industry analysts.
  • Analyze the planning cycle discrepancies between silicon manufacturing (5 years) and agent development (1 week) by reviewing semiconductor supply chain timelines versus AI software release schedules.
  • Review Cisco's acquisitions of Asterix and Whitefield to assess their integration status and role in non-human identity management.
  • Compare AI infrastructure demand patterns with the 5G buildout by analyzing historical telecom infrastructure investment data versus current AI capex trends.
  • Look for counter-arguments to the 'neutral third-party security layer' thesis by reviewing security vendor roadmaps and model provider native security features.
Full transcriptexpand
0:00
Unidentified speaker (speaker_0): But oh my goodness, if I ran out of tokens like it is like not having ammunition when you're in the middle of a war.
0:05
Unidentified speaker (speaker_1): AI demand continues to expand at an unprecedented pace. Google was processing more than 3.2 quadrillion tokens a month by May, seven times as many as a year earlier. To keep up with that demand, the entire technology stack underneath is being rebuilt from compute and networking to power and security.
0:24
Unidentified speaker (speaker_0): Security for the very first time has become the pre prerequisite for relevance.
0:28
Unidentified speaker (speaker_1): G2 Patel is president and chief product officer of Cisco, where he runs the organization behind every one of those layers, from the chips and optics inside AI data centers to the company's networking and security businesses.
0:40
Unidentified speaker (speaker_0): Every part of our business is actually central to every major issue that you see with an AI.
0:45
Unidentified speaker (speaker_1): Last fiscal year, Cisco took in more than $9 billion of AI infrastructure. orders from the largest cloud providers. The last technology rebuild on this scale was the dot-com boom, when Cisco briefly became the world's most valuable company. In this conversation, G2 and I talk about why trust will set the pace of AI adoption.
1:05
Unidentified speaker (speaker_1): What happens when AI writes almost all of a company's code, and what makes this shift different from the dot-com dot com era?
1:14
Unidentified speaker (speaker_2): Hi everyone, this is the TechSurge Deep Tech Podcast presented by Celeste Capital. Each episode, we spotlight issues and voices at the intersection of emerging technologies, company building, and venture investment. I am Sri Ram Biswanathon, founding managing partner at Celesta Capital. If you enjoy TechSurge, TechSearch, now is a
1:34
Unidentified speaker (speaker_2): perfect time to hit the like and subscribe button. And while you're at it, you can leave us a review. on your favorite podcast platform. If you're just discovering us, visit TechSearchPodcast.com to sign up for our newsletter and check out the archive of some very interesting past episodes.
1:51
Unidentified speaker (UNKNOWN): We're
1:51
Unidentified speaker (speaker_3): delighted to welcome Jitu Patel to TechSearch Podcast today. Jeethu is the president and chief product officer of Cisco, where he's responsible for the company's entire product portfolio, technology, networking. security, collaboration, observability businesses, and essentially running all of Cisco's directionally, all the products and all of that,
2:11
Unidentified speaker (speaker_3): including some of the acquisitions that you've done with Splunk and others. And Cisco sits at the center of pretty much the core of all the issues that people are dealing with in artificial intelligence. And I can hardly think of anybody else that is in a poll position to give a perspective on what's what's
2:31
Unidentified speaker (speaker_3): coming and what are some of the challenges and opportunities than uh G2. So J2, it's a great pleasure to have you in
2:36
Unidentified speaker (UNKNOWN): the show. It's an honor to be here. Thank you for
2:37
Unidentified speaker (speaker_0): having me, Shay.
2:38
Unidentified speaker (speaker_3): Thank you so much. Um so why don't we start at a very high level on just macro trends? And you know, obviously we're past this phase of training, or we're in the phase of inference and and there's lots of conversations about, you know, prefill and decode and just the scale of the infrastructure that's going up and the
2:57
Unidentified speaker (speaker_3): constraints that go with it from compute to networking to energy to you know, power to all sorts of things. So just give us a you know, frame this for me as to w you know, how different is this relative to some of the other key inflections that we've gone through, including the ones that Cisco has actually driven in the late nineties?
3:18
Unidentified speaker (speaker_0): Cisco has driven and then also the ones that Cisco has missed. Like we we drove the internet, missed the cloud, right? And one of the things that we learned as a lesson was we're not gonna miss AI.
3:30
Unidentified speaker (UNKNOWN): Uh
3:31
Unidentified speaker (speaker_0): and when you have these kind of tectonic market shifts that happen, like I the the reality is is you you can't fight the mega trend, you have to use it as a tailwind and propel you forward. And that's you know, it was we were very intentional about it from the very beginning. And you're right that at this point in time, every part of our business is
3:51
Unidentified speaker (speaker_0): actually central to every major issue that you see within within AI. Um, you know, if you think about infrastructure, we happen to be right there. If you think about silicon, we happen to be right there. If you think about trust and safety and security, we happen to be right there. Observability, we are there. If you want a big data platform for machine data, we happen to be right
4:11
Unidentified speaker (speaker_0): there. That was not as obvious to people um three, four years ago. And
4:17
Unidentified speaker (UNKNOWN): uh
4:17
Unidentified speaker (speaker_0): part of it was just making sure that you can
4:21
Unidentified speaker (UNKNOWN): a
4:21
Unidentified speaker (speaker_0): uh adjust the portfolio to really be able to help
4:25
Unidentified speaker (UNKNOWN): with
4:25
Unidentified speaker (speaker_0): this movement. Let's just take a step back and say what's happening in the market right
4:29
Unidentified speaker (UNKNOWN): now.
4:30
Unidentified speaker (speaker_0): I think we are moving to uh from chatbots to this phase of agents and physical AI is shortly to follow. I was wrong. I had thought that we would have a second GP GPT like moment, chat GPT like moment with uh Open Claw.
4:48
Unidentified speaker (speaker_0): And the reason I think I was wrong is because the complexity of going out and using an agent with OpenClaw was too high. The cognitive load was too high for people to go out, I have to get a Mac Mini, I have to get a secure container, I have to make sure I get some kind of uh security and safety guardrails. Um, I have to install OpenClaw. Like that it just seemed like it was very difficult to
5:08
Unidentified speaker (speaker_0): do. I think with these personal agents, you're gonna see a massive structural shift that is about to happen. And right now the the take rate on agents is actually slower
5:18
Unidentified speaker (UNKNOWN): than
5:18
Unidentified speaker (speaker_0): what the chatbot take rate was in some ways, in the sense that it's not moving as fast universally. And part of the reason for that is because people don't trust
5:30
Unidentified speaker (speaker_0): giving it the level of agency that you
5:32
Unidentified speaker (UNKNOWN): need
5:32
Unidentified speaker (speaker_0): to give it. And it's
5:33
Unidentified speaker (speaker_3): complex. I mean it's not easy to do.
5:35
Unidentified speaker (UNKNOWN): It's
5:35
Unidentified speaker (speaker_0): complex, and I think most people if you go to different parts of the world which where you're not living in the Silicon Valley bubble, and you ask people um, are you using an agent? Even the people that raise their hands, if you ask them what the use case is, it's very much it's very much of a chatbot kind of use case. Yeah,
5:53
Unidentified speaker (speaker_3): they're making reservations in restaurants. Yeah, most
5:55
Unidentified speaker (UNKNOWN): of the
5:55
Unidentified speaker (speaker_3): time, right? Or
5:56
Unidentified speaker (speaker_0): or not even that. Sometimes they're asking questions and just getting answers, but through an agent rather than to a um, you know, to a chatbot.
6:03
Unidentified speaker (UNKNOWN): I
6:03
Unidentified speaker (speaker_0): think this notion of long horizon tasks
6:07
Unidentified speaker (speaker_0): that um can be done where you can have, you know, near full autonomy and execution of tasks and jobs um with a human either not in the loop or barely in the loop,
6:22
Unidentified speaker (UNKNOWN): yeah,
6:22
Unidentified speaker (speaker_0): is something that is still foreign to people. Yeah. And there's a paradox over there which is like agents are only as useful as the amount of tools you give it access to.
6:32
Unidentified speaker (UNKNOWN): But
6:33
Unidentified speaker (speaker_0): to give it access to the tools,
6:35
Unidentified speaker (UNKNOWN): uh
6:35
Unidentified speaker (speaker_0): you have a trust issue.
6:37
Unidentified speaker (UNKNOWN): Yeah.
6:37
Unidentified speaker (speaker_0): And I think that's the thing that holds it back right now. So like right now, people are kind of there's a level of trepidation saying if I give it access to my email, Mail, if I give it access to my calendar, if I give it access to my credit card, is this gonna go
6:51
Unidentified speaker (UNKNOWN): you
6:51
Unidentified speaker (speaker_0): know, do things that I don't want it to do?
6:53
Unidentified speaker (speaker_1): Yeah.
6:53
Unidentified speaker (UNKNOWN): And
6:53
Unidentified speaker (speaker_0): there's a real risk on that. And I think we have to solve that problem. Once we solve that problem, the
6:59
Unidentified speaker (UNKNOWN): the
6:59
Unidentified speaker (speaker_0): business model shifts that'll happen in almost every company in the world is going to be pretty profound. Um just give you one example. Like if I'm a retailer and
7:13
Unidentified speaker (UNKNOWN): up
7:13
Unidentified speaker (speaker_0): for the past 30 years, I have actually decided to say the way I'm gonna go direct to consumer is I'm gonna have an advertising model and I'm gonna make sure that I show upsell and cross-sell. So if you want to come in and buy a toothbrush, I'm gonna sell you toothpaste, and then maybe some dental floss associated with it, and maybe you should buy
7:33
Unidentified speaker (speaker_0): um an electric toothbrush, and then maybe buy a water pick, and what was going to be a $5 purchase is going to end up being a $200 purchase or more. That has been the model for retailers forever. Right? Now you have an agent that you're going to give agency to. You're going to give it to your credit card. And the agent goes out and shops
7:53
Unidentified speaker (speaker_0): for you. And if you ask it for a toothbrush, it's going to go to 10 places, get a toothbrush, and tell you the cheapest price of the toothbrush, or decide the cheapest price and get the toothbrush. You might have brand. loyalty to the product, but not to the reseller.
8:05
Unidentified speaker (speaker_3): Right.
8:06
Unidentified speaker (speaker_0): And so as a result, your entire calculus has to shift as a retailer to say, what do I need to do to maintain my um my account control for the customer? And how do I figure out a mechanism where, if there's an intermediary, the algorithm that the intermediary takes to go
8:26
Unidentified speaker (speaker_0): out and give to the end consumer is not one that's going to damage me and that I'm where I just become commoditized. Yeah. I think that's a pretty structural shift that's going to happen in almost every industry that's going to be direct to consumer.
8:40
Unidentified speaker (speaker_3): Yeah. And
8:40
Unidentified speaker (speaker_0): I think the direct to consumer is the first start of it. At some point in time, that might also happen with enterprise businesses, as these agents become the ones that actually shop and do the enunciation. initial information gathering from a company, so on and so forth. Right?
8:55
Unidentified speaker (speaker_3): Well, you know, this is this is very interesting. There's so many layers to this that we can peel and we'll we'll get into that. But I want to sort of step back even, you know, more at a macro level. You know, uh circa 2000,
9:05
Unidentified speaker (UNKNOWN): mm-hmm
9:05
Unidentified speaker (speaker_3): uh I think it was April two thousand, which is when And I think the three companies were at the peak of uh
9:14
Unidentified speaker (UNKNOWN): excess.
9:15
Unidentified speaker (speaker_3): of five hundred billion dollars in market cap. Microsoft, Intel,
9:19
Unidentified speaker (UNKNOWN): and
9:20
Unidentified speaker (speaker_3): Cisco. Cisco.
9:22
Unidentified speaker (speaker_3): And interestingly, Cisco at that time, uh you know, I I knew some of the early guys at that time. I think we talked about some of your uh past colleagues, uh, you know, J Shri. J Shri was at Crescendo, and there's another company called Calpana that you know, that Charlie was at. And those two companies, you know, through the acquisition, Cisco became
9:42
Unidentified speaker (speaker_3): leaders in routing and switching. And that really drove the network wave and it drove it to the 2000 peak. And now you're pretty close to that peak.
9:53
Unidentified speaker (UNKNOWN): And
9:54
Unidentified speaker (speaker_3): so the question really is, is it because of the fact that Cisco is being recognized for its significance in the networking space as it relates to AI and the critical need for getting networking right. Is that the reason that you think the market is looking at is this a redo of, you know, Cisco as a networking giant,
10:14
Unidentified speaker (speaker_3): or is it broader than that based on your portfolio of assets, including, you know, Splunk and security and trust and all sorts of things that you're actually adding to your to your bag of goods, if you will? I
10:26
Unidentified speaker (speaker_0): think there's multiple factors. where the sentiment on Cisco changed. The first one was for about
10:32
Unidentified speaker (UNKNOWN): a
10:32
Unidentified speaker (speaker_0): decade to a decade and a half, we had stopped innovating. And in my mind, the formula is very simple: innovation is the lifeblood of a tech company. The moment you stop making products that are amazing, that delight people, the rest of it is just window dressing. You know, like you have to start with a great product.
10:52
Unidentified speaker (UNKNOWN): And
10:52
Unidentified speaker (speaker_0): once you have a great product, everything else really matters. But if you don't have a great product, nothing else matters. So that's the first thing is we were,
11:00
Unidentified speaker (UNKNOWN): we
11:00
Unidentified speaker (speaker_0): had stopped innovating. The second thing had become is like six years ago when I joined, we had become like a holding company.
11:08
Unidentified speaker (speaker_0): 250 acquisitions. You've had like an enormous amount of um people that had joined the company from the outside that wanted to just be GMs. And the mental model of the product organization was in order for me to succeed, I have to be a general manager. And so what had happened was you had a bunch of these fiefdoms that have been
11:28
Unidentified speaker (speaker_0): created culturally,
11:30
Unidentified speaker (UNKNOWN): all
11:30
Unidentified speaker (speaker_0): of them running their own PL,
11:33
Unidentified speaker (speaker_0): but no one thinking about the composite value of Cisco, and all of them feeling starved in their business. businesses, thereby not being able to effectively win against pure play players. And what we did
11:47
Unidentified speaker (speaker_0): when I joined Chuck and I talked about this, and he and we said, hey, look, we we need to move from a um holding company to a platform company. Yeah.
11:57
Unidentified speaker (UNKNOWN): And
11:58
Unidentified speaker (speaker_0): it has to be a platform that's tightly integrated components but loosely coupled, because you can't be arrogant. enough to think that you're gonna go out
12:04
Unidentified speaker (UNKNOWN): and
12:05
Unidentified speaker (speaker_0): um you know have every customer buy everything from you in the full stack. But boy, when they do buy two things from you, they should work like magic. Yeah. You know, and
12:14
Unidentified speaker (UNKNOWN): so
12:14
Unidentified speaker (speaker_0): we did that. And by the way, now we have a single platform
12:19
Unidentified speaker (speaker_0): for our silicon,
12:21
Unidentified speaker (UNKNOWN): we
12:21
Unidentified speaker (speaker_0): have a single platform for our um you know, kind of management plan. plane, which is Cisco Cloud Control. So one one product that manages every product that we ever one platform that manages every product that
12:35
Unidentified speaker (UNKNOWN): we
12:35
Unidentified speaker (speaker_3): have.
12:36
Unidentified speaker (speaker_3): It's
12:36
Unidentified speaker (UNKNOWN): cloud control.
12:37
Unidentified speaker (speaker_3): Yeah. It's
12:37
Unidentified speaker (speaker_0): cloud control. And we have a single data fabric.
12:39
Unidentified speaker (speaker_3): Yeah.
12:39
Unidentified speaker (speaker_0): Right?
12:40
Unidentified speaker (UNKNOWN): And
12:40
Unidentified speaker (speaker_3): you made a lot of acquisitions prior to your time.
12:44
Unidentified speaker (UNKNOWN): I
12:44
Unidentified speaker (speaker_3): guess the question is, if you're in this generation, if you're actually thinking of yourself as much broader platform. um and yet you're a your you know, networking is a very core to what you are
12:55
Unidentified speaker (UNKNOWN): about. That's core.
12:57
Unidentified speaker (speaker_3): Uh but you're still in a picks and shovels sort of a company.
13:00
Unidentified speaker (speaker_0): Yeah.
13:00
Unidentified speaker (UNKNOWN): And
13:00
Unidentified speaker (speaker_3): you run into some interesting challenges, right? I mean you're you know, uh uh just like Intel is r runs into a lot of challenges when its customers are so big that they are not dependent on an ODM or an OEM to bring me a box. They can, you know, they can sort of source contract manufacturing and build their
13:20
Unidentified speaker (speaker_3): own server boxes for their data centers. So do you see a comparable sort of challenge with Cisco where some of your your customers, large customers, hyperscalers, they could actually build their own networking boxes for these large where where do you draw the line where you are they are your customer versus your competition versus your you know
13:40
Unidentified speaker (UNKNOWN): I actually think of
13:40
Unidentified speaker (speaker_0): it as you you have different levels so we serve hyperscalers, neo clouds, sovereign clouds, service providers, and enterprise. Each one of them
13:51
Unidentified speaker (speaker_0): uh participates differently in our stack.
13:53
Unidentified speaker (UNKNOWN): So
13:53
Unidentified speaker (speaker_0): firstly, let's take a step back and say what is the core value we bring in networking for AI workloads? Because that's where the volume is like on the data center for the enterprise as well as uh campus branch for the enterprise, what we had to do was build great products and you had to just make sure that they were better, cheaper, faster, right? And good quality and they were
14:13
Unidentified speaker (speaker_0): integrated with the rest of the business. And the big simplify the management,
14:18
Unidentified speaker (UNKNOWN): make
14:18
Unidentified speaker (speaker_0): sure that you actually have world class uh resilience and ensure that you've tied it into the
14:27
Unidentified speaker (UNKNOWN): infrastructure.
14:27
Unidentified speaker (speaker_0): Uh the inf tied security into the fabric of the switch.
14:31
Unidentified speaker (UNKNOWN): So
14:32
Unidentified speaker (speaker_0): that's the first
14:33
Unidentified speaker (UNKNOWN): thing. But when
14:34
Unidentified speaker (speaker_0): you go to hyperscalers and you go to AI workloads, we had to make a strategic call. This was made before my time.
14:41
Unidentified speaker (UNKNOWN): But
14:41
Unidentified speaker (speaker_0): the strategic call is: do we stay reliant on merchant silicon from third parties, or do we decide to go out and build our own? And we decided we acquired a company called Libra and we built our own merchant silicon
14:57
Unidentified speaker (speaker_0): that we started with um you know a few lines and now we basically have one of the largest programs
15:05
Unidentified speaker (speaker_3): on
15:05
Unidentified speaker (speaker_0): the silicon side. But just the silicon is not enough. In fact, what we have is
15:11
Unidentified speaker (UNKNOWN): our
15:11
Unidentified speaker (speaker_0): own silicon,
15:13
Unidentified speaker (UNKNOWN): but
15:13
Unidentified speaker (speaker_0): we make our own photonics as well. We made an amazing acquisition of a company called Acacia. Yep.
15:18
Unidentified speaker (UNKNOWN): I'm
15:18
Unidentified speaker (speaker_3): aware of it.
15:18
Unidentified speaker (UNKNOWN): And
15:19
Unidentified speaker (speaker_0): Acacia has done amazingly well.
15:21
Unidentified speaker (UNKNOWN): And
15:22
Unidentified speaker (speaker_0): so on the switch side of the house, our our revenue from hyperscalers would be zero if it weren't for
15:30
Unidentified speaker (UNKNOWN): um
15:30
Unidentified speaker (speaker_0): silicon.
15:31
Unidentified speaker (speaker_3): Yep.
15:31
Unidentified speaker (speaker_0): Right? The the network processing. Sure. On the coherent optics and optical transceiver side of the house, that's forty percent of our revenue on hyperscalers, give
15:42
Unidentified speaker (UNKNOWN): or
15:42
Unidentified speaker (speaker_0): take in in a given quarter for um
15:45
Unidentified speaker (UNKNOWN): uh
15:46
Unidentified speaker (speaker_0): we're in the quiet period right now, so I'm not reflecting on
15:48
Unidentified speaker (UNKNOWN): what
15:48
Unidentified speaker (speaker_0): it is for the this quarter, but in general. Um and and so that's the photonic side. One
15:54
Unidentified speaker (speaker_3): of the things that you've actually I've I've heard you speak about this, which I think is a fantastic framing. Uh there's sort of these five forces, you know, agents and inference, tokens, trust and control. I mean,
16:06
Unidentified speaker (UNKNOWN): is
16:06
Unidentified speaker (speaker_3): is that broadly sort of the four big initiatives that you can map every one of your businesses under? And if that's the case, w you know, is there one that actually drives more of the margin than others, or is that all equal or i you know, how do you think about it?
16:21
Unidentified speaker (UNKNOWN): I
16:21
Unidentified speaker (speaker_0): mean, um between all of those forces?
16:23
Unidentified speaker (UNKNOWN): All of the
16:23
Unidentified speaker (speaker_3): yeah, the forces, yeah.
16:24
Unidentified speaker (speaker_0): I don't know if I think of those as individual businesses. Those forces kind of combine together and drive our business forward from a demand perspective. Like we're not a demand constrained business.
16:34
Unidentified speaker (UNKNOWN): But is one
16:35
Unidentified speaker (speaker_3): of them actually gives you a moat relative to competition or
16:39
Unidentified speaker (UNKNOWN): security.
16:39
Unidentified speaker (speaker_0): I think security
16:40
Unidentified speaker (UNKNOWN): is.
16:41
Unidentified speaker (speaker_0): By the way, I don't think we've been re-rated for security yet.
16:43
Unidentified speaker (UNKNOWN): I
16:44
Unidentified speaker (speaker_0): see.
16:44
Unidentified speaker (speaker_3): Yeah.
16:44
Unidentified speaker (speaker_0): And so I think that'll come as the growth rates of security go up and they will go up sure
16:50
Unidentified speaker (UNKNOWN): over
16:50
Unidentified speaker (speaker_0): time.
16:51
Unidentified speaker (speaker_3): Yeah.
16:52
Unidentified speaker (speaker_0): Um because we feel like there's such a tremendous opportunity in security for us
16:58
Unidentified speaker (UNKNOWN): um as
16:58
Unidentified speaker (speaker_0): we go through this. Who
16:59
Unidentified speaker (speaker_3): do you who do you run into in in a competitive landscape?
17:03
Unidentified speaker (speaker_0): Look, I uh on the networking side, on the data center side, a competitor is Arista. On
17:07
Unidentified speaker (UNKNOWN): the
17:07
Unidentified speaker (speaker_0): campus brand side, a competitor competitors is HP J H HP H
17:10
Unidentified speaker (UNKNOWN): B Juniper. Yeah.
17:11
Unidentified speaker (speaker_0): Uh on the security side it's Palo Alto
17:13
Unidentified speaker (UNKNOWN): and
17:13
Unidentified speaker (speaker_0): CloudStrike. You know, security side you also have someone like Microsoft and Google.
17:17
Unidentified speaker (speaker_3): Right.
17:17
Unidentified speaker (speaker_0): Um then you go into um uh the Splunk side of the house, you might have um you know folks like Datadog and Observability. Yeah. You have um you know um you have folks like Google Chronicle. You
17:31
Unidentified speaker (UNKNOWN): have
17:31
Unidentified speaker (speaker_0): Microsoft Sentinel.
17:32
Unidentified speaker (speaker_3): Yeah.
17:33
Unidentified speaker (speaker_0): So I think we have multitude of competitors, but when you actually think about how many companies are full stack. Right. for infrastructure from the silicon and photonics all the way to the top, which is agents, I don't think there's a single competitor that does that.
17:49
Unidentified speaker (UNKNOWN): But you know,
17:49
Unidentified speaker (speaker_3): in the old days, I think as you were thinking about wireless and switches and routers, you had this incredible software architecture iOS around your boxes. That was very sticky. In fact, I I have talked about this in some of the other episodes where Nvidia
18:07
Unidentified speaker (speaker_3): and the CUDA ecosystem is sort of akin to what Cisco did in the old days with iOS. And it became sort of the platform of choice for anybody that was, you know, enterprise grade security, networking, and you know interconnects and all of that. They could write to that and they could extend that network. Do you see that akin to what Nvidia
18:27
Unidentified speaker (speaker_3): has got now, or is it different uh uh from a uh from an AI workload standpoint?
18:33
Unidentified speaker (speaker_0): Um I'd say that what we have right now is that is a strategic control point for us.
18:40
Unidentified speaker (speaker_3): Yeah.
18:41
Unidentified speaker (speaker_0): Uh and we're not at all confused about that.
18:43
Unidentified speaker (speaker_3): Right. But you would weave Leave security in that layer.
18:46
Unidentified speaker (UNKNOWN): In
18:47
Unidentified speaker (speaker_0): that layer. Okay. Yeah. I think that's the thing that actually makes it super valuable.
18:51
Unidentified speaker (speaker_0): Is our sec our networking friends don't happen to have a security stack. Our security friends don't happen to have a networking stack. We happen to have both. Not only do we have both, we have taken security and fused it into the fabric of the network.
19:05
Unidentified speaker (speaker_3): Sure.
19:05
Unidentified speaker (speaker_0): So we have taken a DPU and put it on a switch so that the firewall is not a separate device. We have then started to figure out what we can do to make sure that the enforcement points are everywhere on the fabric. So before you know it, I can have a top or rack switch with an enforcement point
19:22
Unidentified speaker (speaker_3): for
19:23
Unidentified speaker (speaker_0): a mini firewall. Or I can have it on a server, or I can have it as an EVPF agent that's sitting in user space, but still being able to go out and look at every signal and every process that terminates on the host.
19:36
Unidentified speaker (speaker_1): Yeah.
19:37
Unidentified speaker (speaker_0): And that ability to be able to make sure that we are fully hyper distributed in security
19:42
Unidentified speaker (UNKNOWN): and
19:42
Unidentified speaker (speaker_0): baked into the fabric of the network gives us this one advantage that is actually really important. You have to assume that the adversary has already infiltrated your network.
19:52
Unidentified speaker (speaker_1): Yeah.
19:53
Unidentified speaker (UNKNOWN): And
19:53
Unidentified speaker (speaker_0): what you're trying to do is prevent lateral movement. Where does lateral movement happen? On the network. Who has the most amount of um
20:01
Unidentified speaker (speaker_0): um you know kind of capability for lateral
20:03
Unidentified speaker (UNKNOWN): who has the most amount of visibility?
20:05
Unidentified speaker (speaker_0): Cisco does. So our ability to take the telemetry off the network and off security and of observability and fuse them together and correlate
20:15
Unidentified speaker (UNKNOWN): them
20:16
Unidentified speaker (speaker_0): because of the fact that we have Splunk
20:19
Unidentified speaker (speaker_0): is unmatched. And
20:21
Unidentified speaker (UNKNOWN): I think
20:21
Unidentified speaker (speaker_0): most people didn't fully, uh we didn't do a good enough job of fully explaining to folks the cash. calculus behind the acquisition of Splunk, which is super strategic.
20:30
Unidentified speaker (speaker_3): This is interesting because this almost presupposes that the moat that you're creating, I can see it uh as to how that's extremely relevant when you're talking about a frontier model. And you actually have have got this uh you've said something about this uh pizza recipe problem, which I found very interesting, where
20:46
Unidentified speaker (UNKNOWN): you
20:46
Unidentified speaker (speaker_3): said, you know, there's no point in using a $10 ten trillion dollar parameter model to go figure out a recipe for pizzas,
20:53
Unidentified speaker (UNKNOWN): which
20:53
Unidentified speaker (speaker_3): is absolutely right, which is what most people do. Yeah. They don't know what else to do. That's
20:56
Unidentified speaker (speaker_0): right.
20:57
Unidentified speaker (speaker_3): But your argument in that conversation that you had with one of your uh earlier podcasts was that you know there are places where it makes sense for a SLM uh, you know, uh in enterprise.
21:10
Unidentified speaker (speaker_0): Yeah.
21:11
Unidentified speaker (speaker_3): And if you were to think about a within enterprise language model, that is much smaller, very domain specific, does that value add or that moat translate naturally to that as well in terms of how it gets security and the model getting fused?
21:27
Unidentified speaker (speaker_0): It does. And let me tell you how, um, because this is a few things that we've been really working on that I'm super excited
21:33
Unidentified speaker (UNKNOWN): about.
21:35
Unidentified speaker (speaker_0): The ability for an organization to effectively and efficiently use tokens to get tasks done is actually directly proportionate to the security posture. Yeah.
21:46
Unidentified speaker (UNKNOWN): And
21:46
Unidentified speaker (speaker_0): your security posture is directly proportionate to your ability to use agents to transform your business.
21:52
Unidentified speaker (speaker_3): Sure.
21:52
Unidentified speaker (speaker_0): Right? Because if you don't trust the agents, you're not going to delegate work to it. If you
21:55
Unidentified speaker (UNKNOWN): don't
21:55
Unidentified speaker (speaker_0): delegate work to it, you're not going to be a relevant enterprise. So security for the very first time has become the prerequisite. for relevance. Right. And if you think of that and say, okay, so now, given that that is the case, we have to make sure that we have extreme level of
22:15
Unidentified speaker (speaker_0): visibility on the patterns of token usage. And tokenomics Economics is no longer just a budget management exercise, but it is actually a universal registry. Sure. of every agent within your organization and what they're using. And at what point in time are you seeing a behavior pattern
22:35
Unidentified speaker (speaker_0): that is not aligned with what you w thought that it should be doing and therefore be able to pull a kill switch because that agent is going out and excessively consuming tokens Sure. because the worst way that you could actually be put in a compromising position is an adversary decides
22:55
Unidentified speaker (speaker_0): to exhaust your tokens.
22:56
Unidentified speaker (speaker_2): Sure.
22:57
Unidentified speaker (speaker_0): And once you've exhausted your tokens and you don't have enough,
23:00
Unidentified speaker (UNKNOWN): it
23:00
Unidentified speaker (speaker_0): then gives you a set of attacks. And so you have to have a very good token efficient token management play play. And what we've done is we've built the entire stack on Splunk
23:10
Unidentified speaker (speaker_3): for
23:10
Unidentified speaker (speaker_0): tokenomics tokenics that ties into our AI defense stack,
23:14
Unidentified speaker (UNKNOWN): which
23:14
Unidentified speaker (speaker_0): does everything. from algorithmic red teaming of the model all the way to agent behavior monitoring and runtime enforcement or guardrails.
23:24
Unidentified speaker (speaker_3): Yeah.
23:24
Unidentified speaker (speaker_0): That to me, by the way, is like no one does that while still tying into the deep bowels of the infrastructure.
23:32
Unidentified speaker (speaker_3): This is very interesting because you know today, you know, to in in the context of tokenics, tokenomics, everybody sort of thinks about compute. you know, cost, energy per user, and and the metric everybody uses is dollar per token per user per watt.
23:45
Unidentified speaker (speaker_0): Yeah.
23:46
Unidentified speaker (speaker_3): But what you're painting is this incredible need for enterprises to actually think about security in the same vein as, you know, and and the the per user could be per agent.
23:57
Unidentified speaker (UNKNOWN): So
23:57
Unidentified speaker (speaker_3): you almost have to think about security in conjunction with everything else uh for scaling. the tokenomics to really make it, you know, make it viable within the enterprise, right?
24:07
Unidentified speaker (UNKNOWN): I
24:07
Unidentified speaker (speaker_0): I think it's the issue of our time.
24:09
Unidentified speaker (speaker_3): But why is it that that people don't think about that when they define the economics for the frontier model or you know or for inference? They only use it as dollar per token per user per or per agent
24:21
Unidentified speaker (UNKNOWN): per year. I actually
24:22
Unidentified speaker (speaker_0): think um
24:24
Unidentified speaker (UNKNOWN): i
24:24
Unidentified speaker (speaker_0): individual token consumption is actually gonna be a very temporal metric.
24:29
Unidentified speaker (UNKNOWN): As long as it's
24:29
Unidentified speaker (speaker_3): as long as it's an L L M. But if it's an agent it won't be, would it?
24:32
Unidentified speaker (speaker_0): No, even in agents, what you will have is not the consumption of tokens.
24:37
Unidentified speaker (UNKNOWN): What
24:38
Unidentified speaker (speaker_0): you'll have is what is the cost per completion of a task. Well
24:42
Unidentified speaker (speaker_3): that's a different that's an abstracted metric.
24:43
Unidentified speaker (UNKNOWN): It's gonna be an
24:44
Unidentified speaker (speaker_0): abstracted metric. Yeah,
24:45
Unidentified speaker (speaker_3): sure.
24:46
Unidentified speaker (speaker_0): But on the token side, I think the correlation has not happened yet to security posture. Yeah. Because we're too early in the game.
24:52
Unidentified speaker (speaker_3): Yeah.
24:53
Unidentified speaker (speaker_0): And um I don't think most people understand that, oh my goodness, if I ran out of tokens, like it is like not having ammunition when you're in the middle of a war. Sure. Right? Yeah. If you don't have tokens,
25:04
Unidentified speaker (speaker_3): yeah.
25:05
Unidentified speaker (speaker_0): Or if your tokens ran out because you your agents got tricked into running out of tokens.
25:11
Unidentified speaker (UNKNOWN): Uh
25:12
Unidentified speaker (speaker_0): and then all of a sudden your attacks started. What are you going to do? Yeah. And this problem is actually a pretty profoundly big problem that we haven't really crocked our
25:23
Unidentified speaker (speaker_3): heads
25:23
Unidentified speaker (speaker_0): around yet.
25:24
Unidentified speaker (speaker_3): Especially when it comes to the agentic workloads, because I think we're we're in the cut, you're not even in the foot of the curve, as
25:29
Unidentified speaker (UNKNOWN): we speak.
25:30
Unidentified speaker (speaker_0): And you have to assume that the adversaries, if they have the same access to the same technology that we do, your attacks are going to be at machine scale and increasingly. more sophisticated.
25:40
Unidentified speaker (speaker_3): Right.
25:40
Unidentified speaker (speaker_0): Your defenses are going to have to be at machine scale. So there's no way that you can defend against those attacks through just a human workforce. You have
25:47
Unidentified speaker (UNKNOWN): to do it
25:48
Unidentified speaker (speaker_0): through an agent workforce.
25:49
Unidentified speaker (speaker_3): Right. Right.
25:49
Unidentified speaker (UNKNOWN): And
25:50
Unidentified speaker (speaker_0): if your agent workforce does not have the fuel to fuel it,
25:54
Unidentified speaker (UNKNOWN): I
25:54
Unidentified speaker (speaker_0): mean, you're out of business.
25:56
Unidentified speaker (speaker_3): Yeah. When you start thinking about agentique agent workflows and some of these large language models, I am buying these. large language models on the sophistication of the tasks that they can actually complete in the shortest time at the shortest cost and all of that.
26:08
Unidentified speaker (UNKNOWN): That's right.
26:09
Unidentified speaker (speaker_3): At what point in time do you think enterprise users are going to demand some of the things that you guys do, which are so critical for the enterprise as a prerequisite or or a requirement for the model providers to sort of encapsulate in their offering? Or do you see the enterprise users? using
26:24
Unidentified speaker (UNKNOWN): No, I
26:24
Unidentified speaker (speaker_0): don't think at some point in time I don't think you can ask for I think y it is quite evident right now
26:30
Unidentified speaker (UNKNOWN): that
26:31
Unidentified speaker (speaker_0): you're gonna have multiple model providers.
26:33
Unidentified speaker (speaker_3): Sure.
26:33
Unidentified speaker (UNKNOWN): Of
26:33
Unidentified speaker (speaker_0): course it is much harder to go out and institute the things that we're talking about on a per model basis rather than do it across models, across clouds, across agents. And so from that perspective, I don't really think that there'll be a baseline. Think about every SaaS company, every SAS SaaS company increased their
26:53
Unidentified speaker (speaker_0): um their level of safety and security. When Salesforce first came out, they continually keep kept increasing it. But there was a whole new market that was created around CASBY.
27:04
Unidentified speaker (speaker_3): Yep,
27:05
Unidentified speaker (speaker_0): of course. Right? I think this is not going to be that different. Like you'll you will have model providers progressively get better and better.
27:11
Unidentified speaker (speaker_3): Yeah.
27:12
Unidentified speaker (speaker_0): To have um model alignment is issues actually get
27:16
Unidentified speaker (UNKNOWN): you know
27:17
Unidentified speaker (speaker_0): resolved as much as they can. I think mechanistic interpretability where you know why a model does what it does is going to be super important.
27:23
Unidentified speaker (speaker_1): Yeah.
27:24
Unidentified speaker (UNKNOWN): Um
27:24
Unidentified speaker (speaker_0): and I think that science that s field of study is still in its very early days.
27:30
Unidentified speaker (UNKNOWN): Well I
27:30
Unidentified speaker (speaker_3): mean it's
27:30
Unidentified speaker (UNKNOWN): i
27:30
Unidentified speaker (speaker_3): it's a natural progression from chain of thought. It is
27:33
Unidentified speaker (UNKNOWN): right it is
27:34
Unidentified speaker (speaker_3): but that's you know it's not g it's probably
27:36
Unidentified speaker (UNKNOWN): Yeah
27:36
Unidentified speaker (speaker_0): it's a natural progression from chain of thought but there's a big d distance between what the model did and reporting to me and giving me visibility and what it did versus why it chose to do it that way. Right. And that's a big big delta between mech interpretability and chain of thought. All right. So I I don't
27:56
Unidentified speaker (speaker_0): think there will be a time
27:59
Unidentified speaker (UNKNOWN): when
27:59
Unidentified speaker (speaker_0): you will not need
28:01
Unidentified speaker (UNKNOWN): a
28:02
Unidentified speaker (speaker_0): neutral third party layer that goes across every model, every app.
28:08
Unidentified speaker (speaker_3): There will not be.
28:09
Unidentified speaker (speaker_0): I think you will always need
28:11
Unidentified speaker (UNKNOWN): a neutral third
28:12
Unidentified speaker (speaker_0): party layer.
28:12
Unidentified speaker (speaker_3): Got it, got it.
28:13
Unidentified speaker (speaker_0): And and given that, that does not mean the models will not provide what's needed. It
28:18
Unidentified speaker (UNKNOWN): just
28:18
Unidentified speaker (speaker_0): means that the bar will keep getting raised
28:20
Unidentified speaker (UNKNOWN): because
28:20
Unidentified speaker (speaker_0): it is almost a certainty at this point in time that everyone is going to invest in an intelligent routing layer, whether they buy it from open router or build their own.
28:28
Unidentified speaker (speaker_3): Sure.
28:28
Unidentified speaker (speaker_0): They will have open rates models and frontier class models.
28:32
Unidentified speaker (speaker_3): Sure.
28:33
Unidentified speaker (speaker_0): And dynamically, you will actually need to make sure that you route your queries to whatever the model is. And as agent swarms get more prominent, you will need to make sure that for the class of tasks that need to get done, if they're relatively simple and there's a way to go out and be able to direct those simple
28:53
Unidentified speaker (speaker_0): tasks to simpler models that can give you a similar outcome, you would do it for much lower of a cost.
28:59
Unidentified speaker (UNKNOWN): Yeah.
28:59
Unidentified speaker (speaker_3): And
28:59
Unidentified speaker (speaker_0): so I I feel like from that perspective, there's going to be this perpetual need for a security and safety-oriented third party that's neutral provider that needs to be there. And ideally, what we feel is when that neutral third party, this is self-serving. but true, when that neutral third party happens to have that security
29:19
Unidentified speaker (speaker_0): fabric baked into the fabric of the network.
29:23
Unidentified speaker (UNKNOWN): Of
29:23
Unidentified speaker (speaker_3): course it's self-serving, but
29:24
Unidentified speaker (UNKNOWN): it's
29:24
Unidentified speaker (speaker_3): the obvious uh.
29:26
Unidentified speaker (speaker_3): Yeah, yeah. No, no, I think it makes perfect sense. But I guess the question though is it goes back to some of the things that you said earlier on, which is it is so complex, people don't understand it, right? I
29:36
Unidentified speaker (UNKNOWN): think that is our challenge.
29:37
Unidentified speaker (speaker_0): Frank,
29:37
Unidentified speaker (UNKNOWN): that
29:37
Unidentified speaker (speaker_3): is your biggest challenge because when I look at all the models that show up, it's very easy for me to be fooled into thinking that I have to use the latest, greatest, sophisticated model for any task that I use, which can be highly, highly punitive in my token economics. And it's like using a big hammer
29:57
Unidentified speaker (speaker_3): for you know a small job that I want to make happening. happen.
30:00
Unidentified speaker (UNKNOWN): But But here's
30:00
Unidentified speaker (speaker_0): also counterintuitive. So I was talking to someone from Open Open Records. Router yesterday. They said to
30:04
Unidentified speaker (UNKNOWN): me
30:05
Unidentified speaker (speaker_0): something that was really interesting.
30:06
Unidentified speaker (UNKNOWN): One
30:07
Unidentified speaker (speaker_0): of the lowest hanging fruits they have, that's a big big delta between mech interpretability and chain of thought. All right. So I I don't think there will be a time when you will not need a neutral third-party third party layer that goes across every model, every app.
30:26
Unidentified speaker (speaker_3): There
30:27
Unidentified speaker (UNKNOWN): will
30:27
Unidentified speaker (speaker_3): not be. There
30:28
Unidentified speaker (speaker_0): will not I I think you will always need a
30:30
Unidentified speaker (UNKNOWN): neutral
30:30
Unidentified speaker (speaker_0): third party layer.
30:31
Unidentified speaker (speaker_3): Got it, got it.
30:32
Unidentified speaker (speaker_0): And and given that, that does not mean the models will not provide what's needed. It just means that the bar will keep getting raised because it is almost a certainty at this point in time that everyone is going to invest in an intelligent routing layer, whether they buy it from open OpenRouter router or build their own.
30:46
Unidentified speaker (speaker_3): Sure.
30:47
Unidentified speaker (speaker_0): They will have open rates weights models and frontier classes. class models.
30:51
Unidentified speaker (speaker_2): Sure.
30:52
Unidentified speaker (speaker_0): And dynamically, you will actually need to make sure that you route your queries to whatever the model is. And as agent swarms get more prominent, you will need to make sure that for the class of tasks that need to get done, if they're relatively simple and there's a way to go out and be able to direct those simple
31:11
Unidentified speaker (speaker_0): tasks to simpler models that can give you a similar output. out uh outcome, you would do it for much lower of a cost.
31:18
Unidentified speaker (UNKNOWN): Yeah. And
31:18
Unidentified speaker (speaker_0): so I I feel like from that perspective, there's going to be this perpetual need for a security and safety oriented third party that's neutral provider
31:27
Unidentified speaker (UNKNOWN): that
31:27
Unidentified speaker (speaker_0): needs to be there. And ideally, what we feel is when that neutral third party, this is self-serving Self serving, but true, when that neutral third party happens to have that security fabric. baked into the fabric of the network.
31:41
Unidentified speaker (UNKNOWN): Of
31:41
Unidentified speaker (speaker_3): course it's self serving, but
31:43
Unidentified speaker (UNKNOWN): it's
31:43
Unidentified speaker (speaker_3): the obvious
31:45
Unidentified speaker (UNKNOWN): thing to do.
31:45
Unidentified speaker (speaker_3): Yeah, yeah. No, no, I think it makes perfect sense. But I guess the question though is it goes back to some of the things that you said earlier on, which is it is so complex. People don't understand it, right?
31:55
Unidentified speaker (UNKNOWN): I think that is our
31:55
Unidentified speaker (speaker_0): challenge.
31:56
Unidentified speaker (UNKNOWN): That
31:56
Unidentified speaker (speaker_3): is your biggest challenge because uh you know, when I look at all the models that show up, uh it's very easy for me to be fooled into thinking that I have to use the latest, greatest, sophisticated model for any task that I use, which can be highly, highly punitive in my token economics. And it's like using a big hammer
32:16
Unidentified speaker (speaker_3): for you know a small job that I want to make happen.
32:18
Unidentified speaker (UNKNOWN): But here's
32:19
Unidentified speaker (speaker_0): also counterintuitive. So I was talking to someone from Open Open an Router yesterday,
32:22
Unidentified speaker (UNKNOWN): and
32:22
Unidentified speaker (speaker_0): they said to me something that was really interesting
32:25
Unidentified speaker (UNKNOWN): one
32:25
Unidentified speaker (speaker_0): of the lowest hanging fruits. they have right now from a routing technology is just making sure that if people just update it from an old model, yeah,
32:34
Unidentified speaker (speaker_3): let's
32:35
Unidentified speaker (speaker_0): say it's n minus five to a new model,
32:38
Unidentified speaker (speaker_0): your cost will go down.
32:41
Unidentified speaker (UNKNOWN): And
32:41
Unidentified speaker (speaker_0): that is something that most people don't realize. That
32:44
Unidentified speaker (UNKNOWN): a
32:45
Unidentified speaker (speaker_0): model, when it gets to a N minus, the efficiency has increased so much in a model that that's more than of the more more more than. the latter models than the than the former ones. That when people build an agent and they leave the model, whatever it was, and then they keep going out and consuming excessive amounts of tokens, but you've had five versions of models after that.
33:05
Unidentified speaker (UNKNOWN): The
33:05
Unidentified speaker (speaker_0): fifth version is not just better output, it is actually at an infinitely lower cost. It'll change in the next I I think it changes because necessity is a mother of invention. Yeah. You just don't have enough tokens to go out and fulfill the needs of AI. Yeah. That you're gonna have to get infinitely more um more efficient. The problem that you have though is the cost of
33:25
Unidentified speaker (speaker_0): tokens is going down at a rate that is not fast enough commensurate to the demand that's going up. Sure. And you now have
33:33
Unidentified speaker (speaker_0): agents consuming seven times more tokens than humans
33:37
Unidentified speaker (UNKNOWN): in the
33:38
Unidentified speaker (speaker_0): world. In February, they exceeded it. So sit from February to October, November, seven X the increase.
33:45
Unidentified speaker (UNKNOWN): That's
33:45
Unidentified speaker (speaker_3): amazing.
33:46
Unidentified speaker (speaker_0): And so I I just feel like this is going to be a perpetual
33:48
Unidentified speaker (UNKNOWN): state.
33:49
Unidentified speaker (speaker_3): But I think it feels like you guys are perfectly positioned.
33:53
Unidentified speaker (speaker_3): You're so well positioned because of the fact that you have that switching routing history, although at a different use case in the in the enterprise networking space. But effectively you're you're routing workloads, if fent essentially. And if you are able to wrap the security and trust and you know visibility and observability around it, I
34:13
Unidentified speaker (speaker_3): think it becomes a a real plumbing that is inevitable for every enterprise to use.
34:18
Unidentified speaker (speaker_0): This is my fantasy, but I'm not sure if it's actually something that'll
34:21
Unidentified speaker (UNKNOWN): have
34:21
Unidentified speaker (speaker_0): a take rate or not. But imagine if we had
34:25
Unidentified speaker (UNKNOWN): we
34:25
Unidentified speaker (speaker_0): have a tokenomics app. Imagine if that tokenomics app
34:28
Unidentified speaker (UNKNOWN): were
34:29
Unidentified speaker (speaker_0): to give you heuristics that can be used as an input to the intelligent routing because I have a forecast and a plan for for for token budget.
34:39
Unidentified speaker (UNKNOWN): If
34:39
Unidentified speaker (speaker_0): I'm exceeding if my forecast is way out of whack with my plan and I'm in the middle of the month, could there Should be a mechanism I could have to go out and instruct my model router that says, Hey, can you just tone it down a little bit?
34:51
Unidentified speaker (UNKNOWN): This
34:51
Unidentified speaker (speaker_3): is this is fascinating because that leads me to exactly the question I want to ask you. If you put on your CIO hat, for instance, uh you know, if tokens are unit of work, uh I mean you alluded to sort of output is what you should really be measuring. Wouldn't it make more sense to have in that tokenomics, how would you measure output that that
35:11
Unidentified speaker (speaker_3): you can actually assign to a set of tokens that have delivered a particular output, how would you measure it? What would be the metric that you would use for output?
35:20
Unidentified speaker (UNKNOWN): Outcomes
35:21
Unidentified speaker (speaker_3): or output?
35:22
Unidentified speaker (UNKNOWN): I
35:23
Unidentified speaker (speaker_0): think it's cost per task.
35:25
Unidentified speaker (speaker_3): Okay. So it's really uh the
35:27
Unidentified speaker (speaker_0): successfully completed task.
35:29
Unidentified speaker (speaker_3): Yeah.
35:29
Unidentified speaker (speaker_0): Um and what the impact of that task is, is what you have to eventually get to.
35:34
Unidentified speaker (speaker_3): I
35:35
Unidentified speaker (speaker_0): just don't think that the market's mature enough for that yet.
35:37
Unidentified speaker (speaker_3): But is it the same for agentic workload? workloads as it is in the sort of the human workloads?
35:41
Unidentified speaker (UNKNOWN): Yeah,
35:41
Unidentified speaker (speaker_0): because i in the agentic workloads, even more so. Because what would you do in the agentic workload? Go get me
35:47
Unidentified speaker (UNKNOWN): a
35:48
Unidentified speaker (speaker_0): pair of shoes.
35:49
Unidentified speaker (speaker_3): Yeah.
35:49
Unidentified speaker (speaker_0): What is my cost of you buying me a pair of shoes that I wanted successfully?
35:53
Unidentified speaker (speaker_3): Right.
35:54
Unidentified speaker (speaker_0): If you can get that for me with a two dollar surcharge, okay.
35:58
Unidentified speaker (UNKNOWN): If
35:58
Unidentified speaker (speaker_0): you can get it for me for a ten 10 cent surcharge, man, I love you.
36:02
Unidentified speaker (speaker_3): Yeah. Yeah. Interesting. So talk to me about how enterprise uh, you know, uh CIOs should think about it. Should they be thinking about certain critical tasks to have unlimited tokens? And and who is the arbitrator on that? Um
36:18
Unidentified speaker (UNKNOWN): I
36:18
Unidentified speaker (speaker_0): think this, you know, the problem with these kind of articles that happen is they are constructing a headline.
36:25
Unidentified speaker (UNKNOWN): But
36:26
Unidentified speaker (speaker_3): explain the article.
36:27
Unidentified speaker (UNKNOWN): Exactly. Let
36:27
Unidentified speaker (speaker_0): me actually give you the context of how this happened. Because
36:29
Unidentified speaker (UNKNOWN): we
36:30
Unidentified speaker (speaker_0): were a bunch of reporters, I was having dinner with them.
36:32
Unidentified speaker (speaker_1): Yeah.
36:33
Unidentified speaker (UNKNOWN): One
36:33
Unidentified speaker (speaker_0): reporter asked me, said, Hey, so do you think that entry-level jobs are gone? And do you think the companies will stop hiring entry-level people? I'm like, firstly, that's the stupidest thing a company can do is stop hiring entry-level people.
36:45
Unidentified speaker (speaker_0): Because the lack of experience of entry-level people is their biggest asset.
36:49
Unidentified speaker (UNKNOWN): Right. Because when
36:50
Unidentified speaker (speaker_0): what they have is complete lack of know-how or bias on why something works a certain way. And that is gold. So you have to make sure that you actually nurture those people that are coming in, but they have to be matched up with people that have better pattern recognition. And the com combination over there, the people that have
37:10
Unidentified speaker (speaker_0): good pattern recognition, their liability is the inability to unlock. unlearn. The liability of the person who's early in career is their inability to detect patterns effectively enough faster. So
37:21
Unidentified speaker (speaker_3): they've not just had enough water under the bridge, if you will.
37:23
Unidentified speaker (speaker_0): But their big asset is they can ask any question
37:26
Unidentified speaker (UNKNOWN): and
37:27
Unidentified speaker (speaker_0): challenge the status quo. And if these folks have an ability to detect a pattern and apply it to that status quo, Of school. then there's like magic which much happens. Let me give you an example of what I'm doing. I found this
37:39
Unidentified speaker (speaker_0): guy
37:40
Unidentified speaker (UNKNOWN): who
37:40
Unidentified speaker (speaker_0): was a CEO of a company that was buried somewhere down in the organization working on something that was not the most strategic thing. So I'm like, all right, Patty, what else can you do? He's
37:49
Unidentified speaker (UNKNOWN): like,
37:49
Unidentified speaker (speaker_0): Gitoo, I have an idea.
37:51
Unidentified speaker (UNKNOWN): I
37:51
Unidentified speaker (speaker_0): know that you talk about how we need to modernize our code base
37:55
Unidentified speaker (speaker_0): in some areas.
37:57
Unidentified speaker (speaker_0): And you know, one of the things I've publicly stated that I wanted to do was
38:01
Unidentified speaker (UNKNOWN): uh
38:02
Unidentified speaker (speaker_0): by the end of this year, this calendar year, I want to have like half a dozen products that are 100% rhythm with AI. Next next year I want to have 70% of our products 100% rhythm with AI, no human lines of code. And then the year after, I want to make sure that in 24 months, I want to have the fantasy was major line share of the majority,
38:22
Unidentified speaker (speaker_0): 98%, 95%, whatever that number
38:24
Unidentified speaker (UNKNOWN): is,
38:25
Unidentified speaker (speaker_0): should be um
38:26
Unidentified speaker (UNKNOWN): there
38:27
Unidentified speaker (speaker_0): should be no legacy code. We have modernized the code because if we can move the code to memory safe languages like Rust and Go,
38:36
Unidentified speaker (speaker_0): um, and majority of the vulnerabilities happen through money through through memory manipulation, we would have massive, massive levels of benefit to be accreted. Patching is hard and all of those things.
38:48
Unidentified speaker (UNKNOWN): So I
38:49
Unidentified speaker (speaker_0): so he said, Jito, I have an idea.
38:51
Unidentified speaker (speaker_0): We should build some world-class simulations. We should have some amazing evals that are built that can actually
39:00
Unidentified speaker (speaker_0): go out and cover the entirety of the corner case of the use cases.
39:04
Unidentified speaker (UNKNOWN): And
39:05
Unidentified speaker (speaker_0): then we should make sure that we have an agent swarm that just starts to go out and take source code and just
39:10
Unidentified speaker (UNKNOWN): rewrite hill
39:10
Unidentified speaker (speaker_0): climbing and just keep doing rewrites
39:13
Unidentified speaker (speaker_0): and see what happens through the simulator. past the evals, doesn't pass the avals, go back to the simulator and keep doing
39:19
Unidentified speaker (UNKNOWN): that. And
39:20
Unidentified speaker (speaker_0): he's like, I think I can do this with a very small group of people and we can try it out for a certain area. I'm like, go nuts.
39:25
Unidentified speaker (speaker_3): Yeah. Go
39:26
Unidentified speaker (speaker_0): do it.
39:26
Unidentified speaker (speaker_3): And give him an unlimited token budget. The
39:28
Unidentified speaker (speaker_0): unlimited token budget was actually what I didn't want him to do was get overly fixated or constrained, constrained by the token economics. Sure. So I'm like, dude, don't worry about the token budget. Right? Okay.
39:41
Unidentified speaker (speaker_3): That's how it happened.
39:42
Unidentified speaker (UNKNOWN): Just
39:42
Unidentified speaker (speaker_0): go nuts. Of course, if he said I'm gonna go out and spend $18 billion in tokens, I would say no.
39:48
Unidentified speaker (speaker_3): Right.
39:49
Unidentified speaker (speaker_0): But
39:49
Unidentified speaker (UNKNOWN): what
39:50
Unidentified speaker (speaker_0): I when I asked him, I'm like, how much would it take? His thinking was a pretty reasonable amount. Out.
39:57
Unidentified speaker (speaker_1): Yeah.
39:57
Unidentified speaker (speaker_0): You know?
39:58
Unidentified speaker (UNKNOWN): And
39:59
Unidentified speaker (speaker_0): so I'm like, all right, it seems like you have a really good idea. on how to do it. It's a pretty good thesis.
40:05
Unidentified speaker (UNKNOWN): We
40:05
Unidentified speaker (speaker_0): have to prove or disprove it. You've got four weeks. Use as many tokens as you want.
40:10
Unidentified speaker (speaker_3): Sure.
40:10
Unidentified speaker (speaker_0): And show me what you can do in four weeks and prove the model out. If you can prove the model out,
40:15
Unidentified speaker (UNKNOWN): by
40:15
Unidentified speaker (speaker_0): the way, engage with the teams because the goal over here is create that model. And then if that model works and if that code got refactored, give that code to the team so that the team can innovate faster on that base of code.
40:30
Unidentified speaker (speaker_3): But that that poses a very interesting problem, right? When you have a catalyst team like Patty, your guy, who's actually got this interesting idea, he's he's got your blessing to use as many tokens as you can, and he's building it, and then he gives it to your core team that actually incorporates it into your mainstream You know, production code. Now something breaks, you know, who takes accountability
40:49
Unidentified speaker (UNKNOWN): No, no, no.
40:49
Unidentified speaker (speaker_0): So I think the testing process has to be super rigorous.
40:52
Unidentified speaker (UNKNOWN): Well,
40:52
Unidentified speaker (speaker_3): even then, if you sort of project it forward to, let's say, RSI kind of a situation where the code is self healing, self doing whatever, where does
41:02
Unidentified speaker (UNKNOWN): accountability? Here's
41:03
Unidentified speaker (speaker_0): my take on that. At some point in time with AI,
41:08
Unidentified speaker (speaker_0): the provenance Finance has to go back to a human. Human has to take the ultimate accountability
41:14
Unidentified speaker (UNKNOWN): for a
41:14
Unidentified speaker (speaker_0): piece of. code that you have. But you're not going to be able to do that by knowing what every every line of code is.
41:19
Unidentified speaker (UNKNOWN): Because
41:19
Unidentified speaker (speaker_0): it's impossible to do that.
41:20
Unidentified speaker (speaker_3): But is that that does that imply that there is a threshold where it stops? Because today Anthropic claims that the latest version, ninety percent of it, was written by Claude. Yeah,
41:30
Unidentified speaker (UNKNOWN): yeah.
41:31
Unidentified speaker (speaker_3): So is there a is that a threshold where it peaks and then there is a human in the in the
41:36
Unidentified speaker (UNKNOWN): No, I
41:36
Unidentified speaker (speaker_0): I actually believe that you will not be able to in the next twelve to eighteen months, very little code will be written by humans,
41:44
Unidentified speaker (UNKNOWN): if
41:45
Unidentified speaker (speaker_0): at all any. It'll be like a thing that you do to show off to people that you you know how to write code. Like most people won't write code. Sure. However,
41:54
Unidentified speaker (UNKNOWN): and
41:54
Unidentified speaker (speaker_0): by the way, very little code will even be able to be reviewed by humans,
41:58
Unidentified speaker (speaker_2): yeah.
41:58
Unidentified speaker (speaker_0): Reviewable by humans because of the fact that there's so much volume of code. Sure.
42:03
Unidentified speaker (speaker_0): So the only thing you can
42:05
Unidentified speaker (UNKNOWN): do is
42:06
Unidentified speaker (speaker_0): make sure that you have world-class simulators and world-class evals and have the evals cover every corner case and then still take your testing pipeline of unit tests and functional tests and add it to that list. And the combination of simulators, evals, and your testing
42:26
Unidentified speaker (speaker_0): pipeline of the traditional should give you better coverage. And what you're doing is rather than testing what is in each line of code, what you're testing is what the code behaves as when executed.
42:38
Unidentified speaker (speaker_3): Well, that is is brilliant. But that that leads me to exactly the question that I'm sure you've poured a lot of your your brain cells on, which is the hugging face thing. So at some point I mean, you know, I would have to assume that they did all of that in that sandbox environment when the model is, you know, went out of the sandbox. So what broke in your
42:58
Unidentified speaker (speaker_3): mind in that s situation?
42:59
Unidentified speaker (UNKNOWN): Well, I
42:59
Unidentified speaker (speaker_0): th I think like this is where the mech interpretability question you what what happened was the goal was very literally being followed.
43:08
Unidentified speaker (speaker_3): Yes.
43:08
Unidentified speaker (UNKNOWN): And
43:09
Unidentified speaker (speaker_0): so why what you need to have, that's why Cisco is in such a an advantageous position.
43:15
Unidentified speaker (speaker_0): But that's also why this is such an important problem to solve.
43:19
Unidentified speaker (speaker_0): Your agent security and safety apparatus that you have must have a mechanism to dynamically and continuously monitor agent behavior for drift. And when you start seeing drift, it has to be able to make sure that you can intercept it at runtime. So
43:39
Unidentified speaker (speaker_0): the the fact that the agent was looking to escape containment was not the issue. The fact that the looking to escape containment did not get detected as it was trying to escape containment.
43:49
Unidentified speaker (speaker_3): Early enough.
43:50
Unidentified speaker (UNKNOWN): Early enough
43:51
Unidentified speaker (speaker_0): was the issue.
43:52
Unidentified speaker (speaker_3): But is that is that going to get compounded with scaling of agent in you know swarms? Because you know, I can see how you can actually monitor and contain and all of that. This is why
44:02
Unidentified speaker (speaker_0): I talk
44:02
Unidentified speaker (UNKNOWN): about that.
44:02
Unidentified speaker (speaker_0): token efficiency is so important in your in your um in your security strategy.
44:07
Unidentified speaker (speaker_3): Yeah.
44:07
Unidentified speaker (speaker_0): Because it it is just a pure security has become an economics game.
44:12
Unidentified speaker (speaker_3): Yeah.
44:12
Unidentified speaker (speaker_0): It is purely economical at this point. So if you have um
44:18
Unidentified speaker (speaker_0): uh an exponential rise in agent swarms,
44:23
Unidentified speaker (speaker_3): yeah,
44:23
Unidentified speaker (speaker_0): you're gonna going to need to make sure that you have an exponential rise at a much, much more effective cost per token for monitoring the behavior of those agent swarms. Yeah. Otherwise, it doesn't
44:35
Unidentified speaker (UNKNOWN): work. This thing
44:36
Unidentified speaker (speaker_0): doesn't work.
44:36
Unidentified speaker (speaker_3): Yeah. But that leads to a to a very important thing that you touched on early on, which is trust as one of the core pillars of what you're trying to do. So explain the trust architecture that you're trying to implement. Because this the trust the architecture. in this kind of a landscape is very different than
44:56
Unidentified speaker (speaker_3): in 2000, for
44:57
Unidentified speaker (UNKNOWN): instance.
44:58
Unidentified speaker (speaker_3): So talk about that. What how do you think about it?
45:00
Unidentified speaker (speaker_0): Number one, you you cannot govern something that you cannot
45:06
Unidentified speaker (UNKNOWN): see.
45:08
Unidentified speaker (UNKNOWN): And
45:08
Unidentified speaker (speaker_0): number two, you cannot trust something you cannot govern.
45:12
Unidentified speaker (UNKNOWN): And
45:13
Unidentified speaker (speaker_0): you have to know that. So what you have to do. first and foremost is make sure that you get the visibility. So,
45:18
Unidentified speaker (speaker_3): observability is a key.
45:19
Unidentified speaker (speaker_0): Observability is a core part of the fabric of security. I think the ability to distinguish between an ad adversary manipulating an agent to have
45:29
Unidentified speaker (speaker_0): unintended behavior and an agent just literally following your instructions and doing something to exercise, poor judgment, those two things. will be indistinguishable. And so, therefore, the telemetry from security and observability will need to get fused
45:44
Unidentified speaker (UNKNOWN): together.
45:45
Unidentified speaker (speaker_0): And as you fuse that telemetry together, you will see the correlation of data sets over there being superbly potent, right? And so
45:54
Unidentified speaker (speaker_0): first you have to make sure that you get a full inventory. What agents are in my environment? Are they sanctioned? Are they not sanctioned? What are they doing? What are they accessing? Are they allowed to access what they're accessing? Are they not allowed to access what they're accessing? All of that has to get done. And
46:11
Unidentified speaker (speaker_3): even if there's some external data infusion that's happening from outside sources,
46:16
Unidentified speaker (UNKNOWN): all of
46:16
Unidentified speaker (speaker_3): that.
46:17
Unidentified speaker (speaker_0): You have to just monitor the entire environment. Okay. Once you've done that,
46:21
Unidentified speaker (UNKNOWN): you
46:22
Unidentified speaker (speaker_0): have to make sure that you assign an identity to every agent.
46:28
Unidentified speaker (speaker_0): but also every thing within your organization beyond humans.
46:34
Unidentified speaker (speaker_0): Um, we have the thing piece covered with non human identity. We had not had the agent piece covered. We acquired a company called Asterix and we acquired a company called Whitefield, though those gave us those assets to do that.
46:46
Unidentified speaker (speaker_3): Let's switch topics to sort of networking in the data center. Yeah. from a scale up, scale out architecture. You and I are involved with one of the companies that Celeste is an investor in from the early days upscale.
47:09
Unidentified speaker (speaker_3): versus UA Link. Where does Cisco stand in that evolution of the standard for scale up and scale out?
47:14
Unidentified speaker (speaker_0): I think the on on the first thing that I would say is currently where we play, we haven't announced anything on scale up.
47:21
Unidentified speaker (speaker_3): Yeah.
47:22
Unidentified speaker (speaker_0): But currently where we play is very, very heavily on scale out and very heavily on scale across. Okay. Right?
47:29
Unidentified speaker (speaker_3): Scale across you mean in multiple sites, all of that?
47:33
Unidentified speaker (speaker_0): Uh A company that has data centers that are hundreds of kilometers apart. Yeah.
47:37
Unidentified speaker (speaker_3): That's That
47:38
Unidentified speaker (UNKNOWN): what I mean need to get
47:39
Unidentified speaker (speaker_0): you know,
47:40
Unidentified speaker (UNKNOWN): I
47:40
Unidentified speaker (speaker_0): don't want to say coherently clustered because but I I want to say they operate like one logical data.
47:45
Unidentified speaker (UNKNOWN): It's
47:45
Unidentified speaker (speaker_3): a monolithic kind of a look and feel
47:47
Unidentified speaker (UNKNOWN): like that.
47:47
Unidentified speaker (speaker_3): That's
47:47
Unidentified speaker (UNKNOWN): right.
47:47
Unidentified speaker (speaker_3): Yeah, it's called
47:48
Unidentified speaker (speaker_0): and I think what you're now starting to see also is scale up is gonna is becoming role level scale up too.
47:55
Unidentified speaker (speaker_3): Sure.
47:56
Unidentified speaker (speaker_0): Right. And so you will We'll see start to see coherence at a role level rather than a the rack level. But I I think like right now what we are um what we are really, really focused on. Firstly, I I believe that in the long run,
48:09
Unidentified speaker (UNKNOWN): like
48:10
Unidentified speaker (speaker_0): Ethernet becomes a standard that actually ends up winning. But leave that aside for a second and
48:14
Unidentified speaker (UNKNOWN): say,
48:16
Unidentified speaker (speaker_0): what are the reasons that and this is an interesting thing to think about. Like we have, we just launched our G300 G three hundred chip. G300
48:24
Unidentified speaker (UNKNOWN): G
48:24
Unidentified speaker (speaker_0): three hundred chip is our scale-out scale out chip. That scale-out scale out chip has close to a quarter trillion transistors on it and is a hundred and two terabit, you know, kind of um chip that is on a
48:39
Unidentified speaker (UNKNOWN): hundred
48:39
Unidentified speaker (speaker_0): square millimeter package substrate. It's
48:41
Unidentified speaker (speaker_3): on the what the the three three nanometer, two nanometer
48:44
Unidentified speaker (UNKNOWN): um yeah. Three
48:44
Unidentified speaker (speaker_0): th three nanometer. Yeah. Right. And
48:48
Unidentified speaker (speaker_0): and when you start thinking about that. it's a tremendous advancement that you're starting to see with like what's happening on the scale outside. With scale across
48:57
Unidentified speaker (UNKNOWN): what
48:58
Unidentified speaker (speaker_0): we're doing is we've actually also added debuffering technology so that you can make sure that if you have a training run that's going across two data centers and you drop a packet, um, we've buffered the packet packets so that
49:10
Unidentified speaker (UNKNOWN): you
49:10
Unidentified speaker (speaker_0): actually don't have to restart the training run because that could be super expensive.
49:14
Unidentified speaker (UNKNOWN): And
49:14
Unidentified speaker (speaker_0): then what we've done is we've made sure that that technology, which was initially built only for hyperscalers,
49:19
Unidentified speaker (UNKNOWN): I
49:20
Unidentified speaker (speaker_0): think what's going to happen is the enterprises that start gonna, especially the top enterprises within financial services and oil and gas, and pharmaceuticals, and life science,
49:29
Unidentified speaker (UNKNOWN): and
49:29
Unidentified speaker (speaker_0): healthcare, and retail, and service providers, and telcos, will actually be building their own data center footprints out
49:35
Unidentified speaker (UNKNOWN): pretty
49:35
Unidentified speaker (speaker_0): aggressively on the AI side.
49:37
Unidentified speaker (UNKNOWN): Um,
49:38
Unidentified speaker (speaker_0): we need to make sure that when they buy the full stack and we we want to make sure that we make it as turnkey for them as possible.
49:44
Unidentified speaker (speaker_3): Yeah.
49:44
Unidentified speaker (speaker_0): So we have scale out, we have scale across, nothing announced on scale up yet.
49:48
Unidentified speaker (speaker_3): Got it.
49:49
Unidentified speaker (speaker_0): But um right now when people say to me, what are you worried about? Up until now, like my take has always been when you equate this to the prior revolutions that have happened on or the prior platform shifts that have occurred, right?
50:04
Unidentified speaker (UNKNOWN): Yeah.
50:05
Unidentified speaker (speaker_0): Let's say that you had a platform shift with the internet and there was an overbuilding of 5G, and people equate this to, oh, there's an overbuilding of 5G that happened. Yeah, but you have you were waiting at that time for demand to catch up.
50:17
Unidentified speaker (speaker_1): Yeah.
50:18
Unidentified speaker (speaker_0): Over here, anything you make available in supply is instantly consumed. So you don't have a demand problem. What
50:25
Unidentified speaker (UNKNOWN): you
50:25
Unidentified speaker (speaker_0): do have though, yeah.
50:27
Unidentified speaker (speaker_0): is a power available. Your equivalent of dark fiber
50:31
Unidentified speaker (UNKNOWN): is
50:31
Unidentified speaker (speaker_0): a dark GPU. If the power is not made available to turn on the GPU.
50:35
Unidentified speaker (speaker_3): That's right.
50:36
Unidentified speaker (speaker_0): That's number one. And number two is the cost of capital.
50:39
Unidentified speaker (speaker_3): Sure. This is very interesting because it reminds me, and I think you you will relate to this in the old days when I used to work at Intel, Andy Grove used to say uh
50:49
Unidentified speaker (UNKNOWN): the
50:50
Unidentified speaker (speaker_3): loss laws of economic economics uh the laws of physics are the same for every center conductor company. The laws of economics are different. And that makes the difference between someone who is successful and someone who's not. And really what he was he was alluding to is the ability to take large bets in CapEx today to build a fab,
51:09
Unidentified speaker (speaker_3): to manufacture a chip four years out for a market that has not been created, for a demand that has not been identified yet.
51:19
Unidentified speaker (speaker_0): For a m also a market that has a one-week planning cycle where you have to do a six-year planning cycle
51:27
Unidentified speaker (UNKNOWN): for the
51:27
Unidentified speaker (speaker_0): fab.
51:27
Unidentified speaker (speaker_3): So so that that was going to be my question because you have a very unique job because not only are you sort of sitting at the top of this giant company, you're also trying to try tr you to know transform this. company to think like a startup because you've got these large bets that you have to make on silicon, which is as you said, it's a long design
51:47
Unidentified speaker (speaker_3): cycle, long capex commitment before revenue can be uh generated. And you also have to figure out how to, you know, fine-tune your orchestration layer or build the the domain specific model, which is weeks in design time. How does one one sort of bridge that? in
52:03
Unidentified speaker (UNKNOWN): a
52:03
Unidentified speaker (speaker_3): large organization like
52:04
Unidentified speaker (UNKNOWN): Cisco?
52:04
Unidentified speaker (speaker_0): We are um one of the things that we have been very deliberate about, and we have um
52:11
Unidentified speaker (UNKNOWN): you know
52:11
Unidentified speaker (speaker_0): if you look at our entire stack, Silicon and Photonics, five year planning cycle.
52:17
Unidentified speaker (speaker_3): Yeah.
52:18
Unidentified speaker (speaker_0): Our systems business
52:20
Unidentified speaker (speaker_0): is about 24 twenty four to 36 thirty six month planning cycle. Our OS system
52:25
Unidentified speaker (UNKNOWN): is
52:25
Unidentified speaker (speaker_0): a fifteen to
52:28
Unidentified speaker (speaker_3): Yeah,
52:28
Unidentified speaker (speaker_0): exactly twenty month planning cycle. Our platform business is a twelve to eighteen month planning cycle. Our m model business
52:37
Unidentified speaker (UNKNOWN): is
52:38
Unidentified speaker (speaker_0): model business is about a three month to four month planning cycle. Application business is about a couple months planning cycle right now. Agents are about a week planning cycle.
52:48
Unidentified speaker (speaker_3): Yeah.
52:48
Unidentified speaker (UNKNOWN): And
52:48
Unidentified speaker (speaker_0): what we have done is gotten really good
52:51
Unidentified speaker (speaker_0): at actually being able to coordinate across those planning cycles. Which I think is when most people say what's your moat, that's the moat.
53:02
Unidentified speaker (UNKNOWN): It
53:03
Unidentified speaker (speaker_3): could be a moat which prevents competition. It can also be the moat that sort of isolates you from going into certain To markets if you just miss some of these cycles.
53:11
Unidentified speaker (speaker_0): Um
53:12
Unidentified speaker (UNKNOWN): no,
53:12
Unidentified speaker (speaker_0): because like I don't think the market's moving any further faster than us on silicon. The market's not moving any faster than us on hardware, the market's not moving any faster than us on agents. We're the fastest moving in every layer of the stack. But what we understand is the variability of the planning cycles of each one that we have to make sure that we keep into
53:32
Unidentified speaker (speaker_0): account.
53:33
Unidentified speaker (speaker_3): So this this leads me to an interesting question, which is you know, we're gonna get to your own personal journey, but it's i it's illustrative that from where you come from, your personal stories, uh has a lot to do with how you approach uh the culture within Cisco to really be very nimble and think about all of this. So
53:53
Unidentified speaker (speaker_3): what have you sort of learned in the six years that you've been there that has required Cisco to do a fundamental rethink of how they do some of these things? versus how have you sort of thought about things that they've done very well and preserved it and sort of extended it? Is there a bright line between these two things that you think of?
54:13
Unidentified speaker (UNKNOWN): Like
54:13
Unidentified speaker (speaker_0): I think like there's a few like one is I have always
54:19
Unidentified speaker (speaker_0): I've been very public about this. Like I had a rough childhood growing up, and so as a result, um, you know, when you look at some of the things
54:27
Unidentified speaker (UNKNOWN): that
54:28
Unidentified speaker (speaker_0): drive you right now, like I don't have a shortage of hunger. Like I never really stopped being hungry. I think hunger is the hardest thing to thing to teach people. You can't teach hunger.
54:39
Unidentified speaker (speaker_3): Yep.
54:40
Unidentified speaker (speaker_0): So I was lucky enough that I was actually very motivated intrinsically. Not I I have never in my career ever needed someone else to motivate me.
54:50
Unidentified speaker (speaker_3): I have
54:50
Unidentified speaker (speaker_0): been motivated. motivated
54:51
Unidentified speaker (UNKNOWN): by
54:52
Unidentified speaker (speaker_0): myself.
54:53
Unidentified speaker (speaker_3): The curiosity drive was very high. Yeah.
54:55
Unidentified speaker (speaker_0): It's like I just, you know, whatever the next wave is, I want to learn it and become the person that's at the center of that wave, and I will actually take all the time to make sure that that happens. And if it
55:05
Unidentified speaker (UNKNOWN): if
55:05
Unidentified speaker (speaker_0): it doesn't happen, then it's on me. Like I have no one to blame. You know, like that's always been the case.
55:10
Unidentified speaker (UNKNOWN): The
55:11
Unidentified speaker (speaker_0): thing that I've realized in a job like this is
55:15
Unidentified speaker (speaker_0): every once in a while
55:18
Unidentified speaker (speaker_0): you will screw up. When you screw up, you have to make sure that you immediately acknowledge your screw up,
55:27
Unidentified speaker (speaker_0): get in front of the team, assume accountability, tell them it was a mistake, give them context on why that mistake might have happened, not as an excuse,
55:38
Unidentified speaker (UNKNOWN): but just so.
55:39
Unidentified speaker (speaker_0): that they understand that that was not intentional,
55:43
Unidentified speaker (UNKNOWN): or
55:43
Unidentified speaker (speaker_0): that you didn't have enough perspective, and then um do something to correct the mistake and try not to make the same mistake over and over again.
55:52
Unidentified speaker (speaker_3): Is that is that a teachable skill or is that an experiential thing that I think
55:56
Unidentified speaker (UNKNOWN): very
55:57
Unidentified speaker (speaker_3): teachable?
55:58
Unidentified speaker (UNKNOWN): I think
55:58
Unidentified speaker (speaker_0): you just have to have enough humility about the fact that we're all human beings. And I think what ends up happening in these in this world of media is sometimes people are put on a pedestal unnecessarily. Like I have insecurities, I have fears, I have trepidations,
56:18
Unidentified speaker (speaker_0): I sometimes worry about things I shouldn't worry about. The more you can be transparent about that to people,
56:25
Unidentified speaker (UNKNOWN): all
56:26
Unidentified speaker (speaker_0): and also give them context on your motivation and why you move as hard as you do, because it's not something against them.
56:32
Unidentified speaker (UNKNOWN): But
56:33
Unidentified speaker (speaker_0): there are times that you're just wired to go out and move really fast.
56:36
Unidentified speaker (UNKNOWN): Uh,
56:37
Unidentified speaker (speaker_0): it gives people context. Yeah.
56:39
Unidentified speaker (UNKNOWN): And I
56:39
Unidentified speaker (speaker_0): think like context setting, we talk about context a lot in AI models. I don't think we talk about context enough for people.
56:47
Unidentified speaker (speaker_0): And
56:49
Unidentified speaker (speaker_0): incomplete stories make people
56:51
Unidentified speaker (speaker_0): make decisions that can look stupid, but were not stupid when you consider the context they had.
56:57
Unidentified speaker (UNKNOWN): That's
56:57
Unidentified speaker (speaker_3): right, right.
56:58
Unidentified speaker (speaker_0): And so I think like giving taking the time to give context
57:02
Unidentified speaker (UNKNOWN): and
57:03
Unidentified speaker (speaker_0): acknowledging the mistakes when you like I'll I'll give you an example. I won't go into the details, but I I had an epic screw up yesterday.
57:11
Unidentified speaker (speaker_0): Epic, you know stupid mistake on my end.
57:15
Unidentified speaker (speaker_0): One of my team members.
57:17
Unidentified speaker (UNKNOWN): He
57:17
Unidentified speaker (speaker_0): was so upset with me. And it was 100% my fault. 100% my fault. But it was not my fault because I actually felt like I was trying to do I I was my context was entirely different. Yeah. And the way it came across was
57:31
Unidentified speaker (UNKNOWN): very
57:32
Unidentified speaker (speaker_0): different from the way that he interpreted it.
57:33
Unidentified speaker (speaker_3): Yeah.
57:34
Unidentified speaker (speaker_0): If I don't own that, shame on me.
57:38
Unidentified speaker (speaker_0): And and that requires that you surround yourself with people. And in these jobs, what ends up happening is oftentimes people can get very, very, very um kind of complimentary of you, even when you're screwing up, because no one wants to. It's not because they're bad
57:58
Unidentified speaker (UNKNOWN): people.
57:59
Unidentified speaker (speaker_0): Hey, like no one wants to go out and tell someone like, hey, man, you screwed up over here. Like that's just too much of a risk,
58:05
Unidentified speaker (UNKNOWN): right?
58:06
Unidentified speaker (speaker_0): So you have to kind of seek these people out. Like I have a few people on my team
58:12
Unidentified speaker (UNKNOWN): that
58:12
Unidentified speaker (speaker_0): very seldom do they give me compliments, you know. But they know exactly when to bring bring me back up when I'm down. And they know to make sure that I don't get too far up and keep me at at a level head.
58:24
Unidentified speaker (speaker_3): It's That's beautiful. And
58:25
Unidentified speaker (speaker_0): I
58:25
Unidentified speaker (UNKNOWN): think that
58:26
Unidentified speaker (speaker_0): when you find those people, you gotta treasure them.
58:28
Unidentified speaker (UNKNOWN): Yeah.
58:29
Unidentified speaker (speaker_3): And
58:29
Unidentified speaker (speaker_0): you gotta make sure that you do whatever you can do with it.
58:31
Unidentified speaker (UNKNOWN): Yeah.
58:32
Unidentified speaker (speaker_3): It's truly, truly a a skill. It's a learned skill.
58:35
Unidentified speaker (UNKNOWN): That's
58:35
Unidentified speaker (speaker_0): a learned skill.
58:36
Unidentified speaker (speaker_3): Yeah.
58:36
Unidentified speaker (speaker_0): But but it's a it's a skill that's very hard to do. I what I've learned at least is the smarter ones learned it early in life. Uh
58:44
Unidentified speaker (UNKNOWN): like
58:44
Unidentified speaker (speaker_0): it took me a while to pick it up. Yeah.
58:46
Unidentified speaker (UNKNOWN): And
58:47
Unidentified speaker (speaker_0): I'm not I still haven't perfected it. And part of it is like when you go really hard at something, you can sometimes get carried away and not understand that your job is as much to do the right thing as it is to motivate people to make sure that they are doing it because they feel like they want to do it.
59:05
Unidentified speaker (speaker_3): Yeah.
59:05
Unidentified speaker (UNKNOWN): Well, you
59:06
Unidentified speaker (speaker_3): know, Jito, I I I gotta tell you, I mean I as you as I hear you talk, I feel like I'm I'm looking into a mirror because a lot of the things that you say, I can completely resonate with
59:15
Unidentified speaker (UNKNOWN): And and
59:15
Unidentified speaker (speaker_3): and I deal with it all the time. It's a it's a struggle. Like all of us are moving at you know such such fast speeds that sometimes you know you don't give it enough cycle time. You don't
59:25
Unidentified speaker (UNKNOWN): give it enough cycle time.
59:26
Unidentified speaker (speaker_0): But by the way, you have to also have people around you
59:28
Unidentified speaker (speaker_3): to remind you.
59:29
Unidentified speaker (speaker_0): And you have to be able to forgive when mistakes get made. Like expectation out of me for perfection would be an unrealistic expectation
59:38
Unidentified speaker (speaker_3): of anybody.
59:39
Unidentified speaker (speaker_0): of anyone, right? But an expectation of me to be arrogant when I've made a mistake is unexc inexcusable.
59:46
Unidentified speaker (speaker_3): Yeah.
59:46
Unidentified speaker (speaker_0): So I would rather that you actually expect from me that I own a mistake rather than I never make a mistake.
59:54
Unidentified speaker (speaker_3): Yeah.
59:54
Unidentified speaker (UNKNOWN): Well,
59:55
Unidentified speaker (speaker_3): I mean there's so much to uh uh in this fast pace that we live in. I think building transformational businesses or taking businesses like Cisco to a completely different level, it's a it's a huge transition for the company. And it's also an opportunity for our own inner transitions. And it's such a delight to hear you speak of all of that. Uh Jito, this
1:00:15
Unidentified speaker (speaker_3): has been such a wonderful, wonderful conversation.
1:00:16
Unidentified speaker (UNKNOWN): It's a pleasure.
1:00:17
Unidentified speaker (speaker_0): Thank you for having me. And um you know, I'm very delighted that we're partnering together with you folks. And
1:00:22
Unidentified speaker (speaker_3): yeah.
1:00:23
Unidentified speaker (speaker_0): I'm sure this is just the beginning and there's a lot more
1:00:25
Unidentified speaker (UNKNOWN): than we've got to do. Well we've had lots
1:00:25
Unidentified speaker (speaker_3): of history in
1:00:26
Unidentified speaker (UNKNOWN): over over
1:00:27
Unidentified speaker (speaker_3): decades, and
1:00:28
Unidentified speaker (UNKNOWN): yeah,
1:00:28
Unidentified speaker (speaker_3): it's a true joy to have you on this podcast. So thank
1:00:30
Unidentified speaker (UNKNOWN): you very much
1:00:31
Unidentified speaker (speaker_3): for being here. Yeah. Cheers.
1:00:41
Unidentified speaker (speaker_2): Thank you for tuning in to the Text Surge podcast from Celeste Capital. If you enjoy this episode, feel free to share it, subscribe, or leave us a review on your favorite podcast platform. We'll be back every two weeks with more insights and discussions of all things deep tech. Bye for now.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights