Honeywell CTO outlines AI-driven autonomy strategy, citing edge computing and deterministic guardrails for industrial applications
Summary
Mustafa Okur interviews Suresh Venkatarayalu, CTO and President of Forge at Honeywell, regarding the company's pivot from automation to autonomy. Venkatarayalu explains that Honeywell has simplified its portfolio to focus on commercial buildings and industrial process automation, leveraging its Forge platform to integrate AI with deterministic control systems. He describes a three-phase AI strategy involving internal knowledge models, partnerships with Nvidia for open-source models in secure hardware, and hybrid architectures for data sovereignty. The discussion highlights use cases in reducing false alarms in ethylene plants and building management systems, as well as collaborations with Google, LS Electric, and Dangote to expand ontologies and meet cybersecurity mandates.
Episode sections
- 0:18 – 1:47
Introduction and Portfolio Simplification
Mustafa Okur introduces Suresh Venkatarayalu, CTO and President of Forge at Honeywell. Venkatarayalu explains that Honeywell has simplified its portfolio by spinning off aerospace and advanced materials, leaving a pure-play automation focus on commercial buildings and industrial process automation.
- 1:47 – 5:43
From Automation to Autonomy and the Forge Platform
Venkatarayalu describes the strategic pivot from automation to autonomy, defined as automation plus AI. He outlines the 'see, think, act, and learn' framework and positions Honeywell's Forge platform as the vehicle to integrate frontier AI models with their deterministic control systems.
- 5:43 – 8:41
The Moat: Deterministic Guardrails for Probabilistic AI
Venkatarayalu argues that Honeywell's moat lies in fusing probabilistic AI models with deterministic control models. The deterministic layer acts as a safety guardrail for mission-critical infrastructure, ensuring safety and reliability that pure AI models cannot guarantee alone.
- 8:41 – 12:25
Three-Phase AI Strategy: Internal Tools, Physical AI, and Hybrid Models
Venkatarayalu details a three-phase approach: Phase 1 involved training internal models on company knowledge for maintenance and engineering. Phase 2 involves partnering with Nvidia to deploy open-source models like Nemotron and Cosmos Reasoning within a secure 'cognition box' to protect IP. Phase 3 anticipates a hybrid architecture of frontier and open-source models to meet regional sovereignty needs.
- 12:25 – 14:24
Edge Computing and Data Sovereignty
The discussion covers the necessity of edge computing for autonomy due to latency requirements (20-60 milliseconds), data transaction costs, and customer demands for data protection. This has led to the development of 'Forge Cognition,' an appliance that sits on top of cloud and control systems.
- 14:24 – 18:52
Hardware Agnosticism and Protocol Challenges
Venkatarayalu states that the intelligence layer is hardware-agnostic, leveraging standard protocols like BACnet, Modbus, and OPC. He identifies challenges in PLC integration and data sharing, noting that adoption requires CXO-level engagement and a shift toward virtualized PLCs.
- 18:52 – 21:51
Use Cases: Reducing False Alarms in Process and Building Management
Two specific examples are provided: a Middle East ethylene plant where AI reduced 60-70% of 100,000 monthly false alarms, and a building management system where AI eliminated 85-95% of false alarms from security and BMS systems, allowing operators to focus on higher-value tasks.
- 21:51 – 35:42
Autonomous Building Agents and ROI
Venkatarayalu outlines five standard agents for autonomous buildings: energy, operational efficiency, comfort, cyber protection, and security. He highlights significant ROI potential, citing examples where AI optimized heating and cooling set points in buildings designed for static comfort, reducing energy waste.
- 35:42 – 36:39
Over-the-Air Updates and Cybersecurity Strategy
Speaker_1 states that industrial OT systems will receive quarterly software updates rather than weekly, citing experiments starting in Q1. Mustafa Okur highlights the growing importance of cybersecurity following recent attacks on European retail supermarkets.
- 36:39 – 37:56
Google Partnership and Air-Gapped Architecture
Speaker_1 explains that Honeywell partnered with Google to integrate Google Data Center (GDC) into the Bill of Materials for Forge Cognition, adopting an air-gapped approach to meet federal defense and government agency security mandates.
- 37:56 – 38:50
Human-in-the-Loop Control Philosophy
Speaker_1 emphasizes that while AI introduces probabilistic elements, deterministic controls remain paramount. The system keeps operators in the loop to improve efficiency (from 10 to 100-200 actions per day) without eliminating human oversight in safety-critical environments.
- 38:50 – 40:01
Codifying Domain Knowledge and Asset Ontologies
Mustafa Okur asks about codifying Honeywell's century of knowledge. Speaker_1 details Phase One, which involves contextualizing 130 asset types across building management, security, and fire systems, deployed at scale in pilots.
- 40:01 – 42:34
Expanding Ontologies via Partnerships (LS Electric and Data Centers)
Speaker_1 describes expanding the ontology by partnering with LS Electric to add 40-48 electrical asset categories. Further expansion into data centers requires integrating liquid cooling, UPS, and IT racks, necessitating a consortium of OEM partners.
- 42:34 – 51:37
Dangote Process Plant Implementation with Nvidia and Dell
Speaker_1 announces a two-week-old press release regarding a deployment with Dangote in Nigeria, involving Honeywell, Nvidia, and Dell. The project aims to codify complex process chemistry and catalysts, offering significantly higher ROI potential than commercial buildings.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| NVIDIANVDA | equity | neutral | So we partnered with Nvidia. We absolutely brought their models like NemoTron, Cosmos Reasoning, and which goes inside our box because our customers demanded the sovereignty of data, model, and an |
| TeslaTSLA | equity | neutral | I foresee it's not going to be like an Apple every two weeks or every week, like Tesla, you deploy software over the air, but at least once a quarter, your ability to upgrade with patches for OT cybersecurity control model changes or the learnings between the different buildings. |
| AlphabetGOOGL | equity | neutral | Uh we partnered with Google very early and we decided that we'll go with Air Gapped approach. So GDC from Google is the one who will be as part of our Bill of Materials spec in Forge Cognition. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
Honeywell spun off Honeywell Aerospace last June and Solvay's Advanced Materials late last year, leaving a pure-play automation portfolio.
Unidentified speaker (speaker_1) ·1:47
How to verify: Check Honeywell's SEC filings (Form 10-K or 8-K) for the completion dates of the aerospace spin-off and the advanced materials divestiture. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceHoneywell has been in business for 145 years.
Unidentified speaker (speaker_1) ·1:47
How to verify: Verify Honeywell's founding date (1885) against the transcript date (2026) or check the company's official history. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell's deterministic models assure 99.6% reliability.
Unidentified speaker (speaker_1) ·6:06
How to verify: Check Honeywell's technical documentation or investor presentations for reliability metrics of their control systems. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Probabilistic AI models like Gemini, GPT, and Anthropic have 85-90% accuracy.
Unidentified speaker (speaker_1) ·7:26
How to verify: Review public benchmarks or performance reports from Google, OpenAI, and Anthropic regarding their model accuracy rates. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell partnered with Nvidia to bring Nemotron and Cosmos Reasoning models inside their hardware.
Unidentified speaker (speaker_1) ·11:10
How to verify: Check press releases from Nvidia and Honeywell regarding their partnership and the specific models deployed. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceAutonomous decisions require latency of 20 to 60 milliseconds.
Unidentified speaker (speaker_1) ·12:59
How to verify: Review technical specifications for industrial control systems or autonomous building management systems regarding real-time processing requirements. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
A Middle East ethylene plant experienced 100,000 false alarms every month, reduced by 60-70% with AI.
Unidentified speaker (speaker_1) ·19:55
How to verify: Check Honeywell press releases from approximately six months prior to the interview (around May 2026) regarding the Middle East plant deployment. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
A building management system produced 250,000 false alarms every quarter, with AI eliminating 85-95%.
Unidentified speaker (speaker_1) ·20:27
How to verify: Check Honeywell case studies or press releases detailing building automation performance metrics. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell completed a 90-day challenge to deploy Forge Cognition at Nvidia's Voyager headquarters.
Unidentified speaker (speaker_1) ·28:57
How to verify: Check press releases from Nvidia and Honeywell around the time of the investor day (likely late 2026 or early 2026) announcing the Voyager deployment. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceHoneywell is working with 15 lighthouse customers across hospitality, data centers, and stadiums.
Unidentified speaker (speaker_1) ·30:21
How to verify: Check Honeywell's investor presentations or press releases for details on their pilot customer base and deployment scale. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell started experimenting with quarterly over-the-air updates for OT systems in Q1.
Unidentified speaker (speaker_1) ·36:06
How to verify: Check Honeywell's Q1 2026 earnings call transcript or press releases for mentions of Q1 2026 pilot programs for OT updates. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell partnered with Google to include Google Data Center (GDC) in the Bill of Materials for Forge Cognition.
Unidentified speaker (speaker_1) ·36:59
How to verify: Search for press releases or partnership announcements between Honeywell and Google regarding Forge Cognition or air-gapped security solutions. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell has codified and deployed 130 asset types across building management, security, and fire systems in its Phase One AI ontology.
Unidentified speaker (speaker_1) ·39:42
How to verify: Review Honeywell's technical white papers or investor presentations detailing the scope of Forge Cognition's asset coverage. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell partnered with LS Electric a year ago to add 40-48 electrical asset categories.
Unidentified speaker (speaker_1) ·40:21
How to verify: Search for the Honeywell-LS Electric partnership announcement dated approximately one year prior to the transcript (Nov 2026) and verify the asset count in press materials. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Honeywell, Nvidia, and Dell announced a deployment with Dangote in Nigeria two weeks prior to the recording.
Unidentified speaker (speaker_1) ·42:34
How to verify: Search for press releases from Honeywell, Nvidia, or Dangote Group dated mid-October 2026 regarding a joint deployment in Nigeria. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceNvidia's Feynman architecture is scheduled to launch in 2028.
Unidentified speaker (speaker_1) ·44:40
How to verify: Check Nvidia's official product roadmap or recent earnings calls for the launch timeline of the Feynman architecture. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceEquinix is a strategic customer for Honeywell in the data center vertical.
Unidentified speaker (speaker_1) ·46:05
How to verify: Review Honeywell's 10-K or 10-Q filings for major customer disclosures or Equinix's vendor lists. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceHoneywell is redesigning its controls architecture to align with Nvidia's Vera Rubin and Ultra architectures.
Unidentified speaker (speaker_1) ·44:40
How to verify: Search for technical briefings or partnership announcements between Honeywell and Nvidia regarding Vera Rubin and Ultra compatibility. For recent events, verify 2026-08-16 through 2026-11-16. Use another period only when the speaker explicitly names it.
Primary sourceDo your own research
- Verify Honeywell's Q3 2026 earnings call transcript for details on the 'Forge' platform revenue contribution and margin impact.
- Check Nvidia's investor presentations for the status of 'Nemotron' and 'Cosmos Reasoning' model availability and licensing terms.
- Review Honeywell's Form 10-K for risks related to the spin-off of aerospace and advanced materials segments.
- Search for the press release regarding the Dangote process plant deployment in Nigeria to verify the consortium partners and project timeline.
- Examine Google Cloud's enterprise security certifications to confirm the 'air-gapped' architecture claims for Forge Cognition.
- Look for independent case studies or white papers from LS Electric regarding the integration of 40-48 electrical asset categories into Honeywell's ontology.
- Compare Honeywell's quarterly update cadence against industry standards for industrial OT cybersecurity patching cycles.
- Investigate Equinix's uptime reports and Honeywell's service level agreements (SLAs) for data center operations to validate the 99.9999% reliability claims.
Full transcriptexpand
- 0:18
- Unidentified speaker (UNKNOWN): Hi
- 0:18
- Mustafa Okur: everyone, welcome to another episode of Tech Disruptors. My name is Mustafa Okur. I'm a senior industrial analyst at Bloomberg Intelligence. BI is part of Bloomberg's research department with 500 analysts and strategists working across all major markets. My focus is covering US manufacturing and the
- 0:37
- Mustafa Okur: conglomerates and electrical equipment makers in that ecosystem. Today we will dive into the automation of US factories, data center design and construction, as well as how a large industrial company, US uses technology and artificial or superintelligence internally. Joining us for that discussion will be Suresh
- 0:57
- Mustafa Okur: Venkatoriolo, the Chief Technology Officer of Honeywell Technologies
- 1:02
- Unidentified speaker (UNKNOWN): and
- 1:02
- Mustafa Okur: the president of its Forge Software Platform. Suresh's bio says he joined Honeywell Technologies in 1995 and has held senior engineering, technology, and business leadership roles, which I interpret as he has seen it all. Suresh, welcome to Tech Disruptors.
- 1:18
- Unidentified speaker (speaker_1): Great to be here, Musifa.
- 1:19
- Mustafa Okur: Lovely. Thank you for joining us. Um Honeywell, Honeywell Technologies, is a US industrial giant that has remade its portfolio recently by spinning off its aerospace business and divesting some non-core assets. The company is now doubling down on its role of automating buildings. For listeners
- 1:39
- Mustafa Okur: who may not be following Honeywell closely, what is your role today and describe what is this new Honeywell Forge?
- 1:47
- Unidentified speaker (speaker_1): Yeah, yeah. Mr. Fa maybe uh a context, Honeywell been there for 145 years and fundamentally it was a controls company, control systems company serving different end markets. And what we really did recently was simplifying our portfolios. So we have spun
- 2:07
- Unidentified speaker (speaker_1): Honeywell Aerospace recently, last June, and then we had the advanced materials, which is SolSty's Advanced Materials late last year. What's Remind Co. is Honeywell Automation, Pure Play Automation Portfolio. That's Honeywell Technologies right now. And what's good about this uh simplification.
- 2:27
- Unidentified speaker (speaker_1): of portfolio, it gives us a focus around commercial building, industrial, process automation, and technology together. As an automation company, and our focus had been in driving efficiency across a multi-uh-industry portfolio. My role as a chief
- 2:46
- Unidentified speaker (speaker_1): technology officer after spending all my 30 years is to really draw. drive one major pivot, which is we say automation to autonomy. And for me, autonomy means automation plus plus, meaning there are a number of different efficiency levers we can drive to improve the efficiency across our
- 3:06
- Unidentified speaker (speaker_1): customer segments moving forward. And that's pretty much where the focus is going to be. Being a controls company, I think we do three things very, very well. We have our install base, our assets, we connect through our sensors and actuators and through our controls platform. And what
- 3:26
- Unidentified speaker (speaker_1): we have delivered for the industry, we we automated our physical world for the last hundred odd years. And now we're asking the question, how do we really bring in this new age frontier models and AI? And how do you really br drive uh the concept called uh think, act, and learn concept on
- 3:46
- Unidentified speaker (speaker_1): top of our controls platform? So that's uh exciting journey ahead for us and uh Forge is going to play an important role.
- 3:53
- Mustafa Okur: Yeah. And um, you know, think and learn was already a thing in the manufacturing base or the buildings uh installed base. for a while. I guess ACT is the new thing here.
- 4:06
- Unidentified speaker (UNKNOWN): Uh
- 4:06
- Mustafa Okur: no,
- 4:07
- Unidentified speaker (UNKNOWN): I
- 4:07
- Unidentified speaker (speaker_1): would actually reverse that
- 4:08
- Unidentified speaker (UNKNOWN): because
- 4:09
- Unidentified speaker (speaker_1): we had C and ACT, meaning
- 4:12
- Unidentified speaker (UNKNOWN): we
- 4:12
- Unidentified speaker (speaker_1): had millions of sensor data and actuators. Operators were using the controls platform to take decisions. And now with the new age AI model, we believe that we have more intelligence added. around the think a side. And then on top of it, we are saying, can we have
- 4:32
- Unidentified speaker (speaker_1): a learning system? So we see this framework to be see, think, act, and learn framework. That's the foundation for the new autonomy architecture moving forward.
- 4:42
- Mustafa Okur: I see. That makes sense. I was thinking more in the sense of, you know, predictive maintenance has been around for a while, but this is a lot more in-depth
- 4:49
- Unidentified speaker (UNKNOWN): and capable.
- 4:50
- Unidentified speaker (speaker_1): In-depth ability, think about the future system where software predicts the asset performance and predicts the downtime or end of life. And it has an ability to go and create a service ticket and drives the whole workflow to closure. I think that's a brilliant
- 5:10
- Unidentified speaker (speaker_1): future compared to what we had in the past. Yes, great deterministic model. but you had a service tickets you need to create and human was in the loop. So those are the possibilities we have.
- 5:21
- Mustafa Okur: Great. End-to-end capability unlocked. If AI capabilities keep improving, but we avoid sort of the extreme outcomes of either extreme utopia or doomsday scenarios. What remains defensible around Honeywell's current office? uh offerings? What is the moat
- 5:40
- Mustafa Okur: around Forge?
- 5:43
- Unidentified speaker (UNKNOWN): You know,
- 5:43
- Unidentified speaker (speaker_1): Mr. Fa, the the way to think about,
- 5:46
- Unidentified speaker (speaker_1): there are a lot of a lot of external stakeholders who asked me, uh, define AI and how do you use AI? And for starters, as a control system company, we always leverage deterministic models. And then if you really go back, that's what we have done the best.
- 6:06
- Unidentified speaker (speaker_1): Machine learning code, deterministic model, using heuristic decision trees. We build this model for our industry, which assured 99.69, because that's what our industry cared for. That's what we are known for. So for the starters, we have been using deterministic model. What is
- 6:26
- Unidentified speaker (speaker_1): has happened in the industry in the last two to three years? What the audience will hear more about like things like Gemini and GPT and anthropic models, they are probabilistic models. They've come into the space. We believe there's a brilliant opportunity for a company like Honeywell to fuse both deterministic
- 6:46
- Unidentified speaker (speaker_1): and probabilistic. Deterministic delivered what I call as C see and act. And uh and it guided the operator to take some specific decision based on a very clear control loss that we have designed and which is which was our IP. Now you have the probabilistic model on the top, which is also learning and
- 7:06
- Unidentified speaker (speaker_1): evolving over a period in time. It's also guiding with reasoning inside. But for us, the MOTIS Mot is are still our controls platform. is our MOT because that assures safety and mission critical infrastructure for our customers. And we believe that that's going to be a guardrail. On top of meaning,
- 7:26
- Unidentified speaker (speaker_1): the probabilistic model, the accuracy could be 85-90%. Yeah. But the guardrail is our deterministic model. So we actually believe our forge architecture will infuse in the physical AI world.
- 7:39
- Unidentified speaker (speaker_1): we still have the deterministic models in place with our controls architecture. And then Forge brings in on top the probabilistic model and still have a guardrail that assures our the core domain expertise plus the mission criticality, what our customers really caring kidding for.
- 7:58
- Mustafa Okur: Yeah, so that is a very unique offering, I guess, from a company like Honeywell. where you have the guardrails or or like a a a sentient second layer, yeah, which is more deterministic all from your history looking over the AI model.
- 8:15
- Mustafa Okur: Because off the shelf you can go to a frontier model company like Anthropic or chat uh OpenAI or whoever or Google
- 8:22
- Unidentified speaker (UNKNOWN): and
- 8:23
- Mustafa Okur: your partners with some of these companies as well. Could you dive into
- 8:28
- Mustafa Okur: um a little bit just about what you talked about? What has Honeywell built in AI that is differentiated in terms of intelligence? And here I'm thinking a little bit about cognition perhaps. And where are you better off using someone else's model?
- 8:41
- Unidentified speaker (speaker_1): I'll probably break into three phases. Uh two years ago, when you actually saw the frontier models um evolved and started delivering the accuracy level to an extent, we leveraged probably two or three core frontier models. We
- 9:01
- Unidentified speaker (speaker_1): leveraged it where we picked our some of our core uh in meaning implicit and explicit knowledge in the company. We trained all of them. So the first generation was taking like a technical publication, CAD drawings, PNID documents, taking service ticket, and
- 9:21
- Unidentified speaker (speaker_1): we built a model called like a maintenance assist. We said, how can you liberate your service technicians who can actually use your model so that you're actually addressing any of the knowledge deficit you may actually have to address the field level service needs? Second, we took the
- 9:41
- Unidentified speaker (speaker_1): same knowledge repository, we gave it to our project engineering field engineering resources because you have a high level of variability impacting your negative EAC or cost of power quality. So we deployed at scale for the last probably 18 to 24 months. The feedback has been phenomenal because again,
- 10:01
- Unidentified speaker (speaker_1): you're addressing a skill deficit.
- 10:03
- Unidentified speaker (speaker_1): in a world where you have uh always a variability in your performances.
- 10:08
- Mustafa Okur: And is this internal?
- 10:09
- Unidentified speaker (speaker_1): It's internal so
- 10:10
- Unidentified speaker (UNKNOWN): far.
- 10:11
- Unidentified speaker (speaker_1): That's a phase one. Phase two, once we got to the level where we had enough data, information, and knowledge that was trained into this model, we said, how do we go back and actually build this whole intelligent layer on top of Forge for us at that point in time, this whole idea of physical AI evolved.
- 10:31
- Unidentified speaker (speaker_1): And during the investors' day, we showcased Honeywell was better positioned because we automated the physical world for all the 145 years. And we said we are ready enough to go and drive the physical AI architecture. That's the reason we kind of started talking about this new layer called intelligence layer with on.
- 10:50
- Unidentified speaker (speaker_1): ontology as a concept, which we'll talk a little bit more in depth. At that point in time, we said the important aspect of the next generation is going to be contextualizing your physical world and physical assets. So we took a different route in taking an open source uh AI
- 11:10
- Unidentified speaker (speaker_1): models because you've you got to really protect the IP and more MOAT that. you talked about. So we partnered with Nvidia. We absolutely brought their models like NemoTron, Cosmos Reasoning, and which goes inside our box because our customers demanded the sovereignty of data, model, and an
- 11:30
- Unidentified speaker (speaker_1): IP. And that's an approach we thought will come pretty nicely. So that's the second generation shift in driving autonomy, because autonomy needs to you to really protect the IP around how you really drive your infrastructure. So that's the phase one and a phase two. In terms of phase three, which we believe the future,
- 11:50
- Unidentified speaker (speaker_1): we have to be ready for both hybrid model, because you may actually have a different model which is ready to solve different use cases at the same time, region by region, country. by country, there may be more and more new sovereignty requirement and specification, which I think we we believe
- 12:10
- Unidentified speaker (speaker_1): we will continue to deal with a hybrid architecture model. We will have both frontier models, open source models, and within an open source, we may pick and choose based on the performance and accuracy and serving and supporting the regional sovereignty needs.
- 12:25
- Mustafa Okur: Yeah, and hybrid in this context means both your models, open models, and frontier models, but also perhaps, and correct me if this is wrong, having AI on the edge on-prem versus having it in the cloud. Is that the right way to think about it?
- 12:39
- Unidentified speaker (speaker_1): I think so, because the reason we are saying autonomy needs to be driven within the four walls of an enterprise for three reasons, Mustafa. One, um, latency matters. when you drive autonomy. Decisions need to be taken in 20 to 60 milliseconds. Uh latency matters. Second, uh the cost
- 12:59
- Unidentified speaker (speaker_1): of data. If you expect everything to be uh driven by cloud, somebody needs to pay for the cost of transacting data to the cloud. Third, I think the customers were really demanding that the data is protected. And I think from all the three, you distribute your compute strategy or intelligence strategy between
- 13:19
- Unidentified speaker (speaker_1): what can you do offline in cloud and what can you do at the edge or within the enterprise four walls. And that's what we are actually evolving, where we said forge has two chapters. Chapter one was we approached with an IoT cloud strategy. Phase two is a more of a appliance we call forge cognition. that has
- 13:39
- Unidentified speaker (speaker_1): this intelligence layer which sits on top of cloud, which sits on top of any control system architecture that drives an autonomy strategy and a physical AI strategy.
- 13:50
- Mustafa Okur: Makes sense. And let's uh keep talking about the application part a little bit more in the in the factory
- 13:56
- Unidentified speaker (UNKNOWN): setting.
- 13:56
- Mustafa Okur: So when people think about automation, so old school automation, let's call it,
- 14:01
- Unidentified speaker (UNKNOWN): they
- 14:01
- Mustafa Okur: think about programmable logic control controllers, uh motion control, robotics maybe. Honeywell today is very strong in sensing and measurement rather than sort of these old school discrete control uh hardware, let's call it. Could you tell us where does Honeywell sit in a truly autonomous factory today? And uh would you be interested in more
- 14:21
- Mustafa Okur: hardware as well as software going to going into the future?
- 14:24
- Unidentified speaker (speaker_1): I think there is an interest in Honeywell. to expand our PLC strategy.
- 14:31
- Unidentified speaker (speaker_1): There is an interest in certain new portfolios like discrete manufacturing, so that we continue to evolve both organic and inorganic, that discussion continues to happen. What I'm pretty excited about this new age intelligence layer that we are trying to build with Forge Cognition is your ability to create an agnostic.
- 14:51
- Unidentified speaker (speaker_1): layer on top of any hardware, any controls platform, as long as customer gives you an access for data access rights, you should be able to really deliver a new uh new value that never existed before. The intelligence layer comes
- 15:10
- Unidentified speaker (speaker_1): from things that are open today in the in in the industry. In buildings world, you can use BACNET protocol and Modbus protocol that are available to extract data to deliver value. And then in the classical process industry world, OPC. So as long as your
- 15:30
- Unidentified speaker (speaker_1): protocol is understood across the board and then you have a customer's willingness to share the data, I think your value of delivering this intelligence layer to deliver into an enterprise level outcome. I think that's the that's the new new way of delivering a value. And that way you can decouple. The key point is:
- 15:50
- Unidentified speaker (speaker_1): do you still understand the domain depth of the industry and do you understand the business context of the industry to deliver value?
- 15:56
- Unidentified speaker (UNKNOWN): I
- 15:57
- Unidentified speaker (speaker_1): think that's the muscle muzzle that all companies are going to be building moving forward.
- 16:00
- Mustafa Okur: That makes sense. And are you seeing is
- 16:02
- Unidentified speaker (UNKNOWN): the
- 16:02
- Mustafa Okur: challenge?
- 16:03
- Mustafa Okur: here more customers may not be willing to share all of their data with you? Or is it the case that sometimes you have to deal with these closed protocols? I mean, PLCs are sort of infamous for being closed platforms, you know, like usually it's difficult to integrate your software into the PLC from, I don't know, Rockwell or XM.
- 16:23
- Mustafa Okur: It's not necessarily easy, right? So is what are the challenges? you're seeing when you sort of work with clients? Is it more on the sort of them saying, hmm, maybe we shouldn't be sharing all this data? Or is it are there sort of hardware-specific uh challenges too?
- 16:38
- Unidentified speaker (speaker_1): I see two types of challenges, Mr. One, PLC is was not designed to expose its protocol.
- 16:47
- Unidentified speaker (UNKNOWN): But
- 16:47
- Unidentified speaker (speaker_1): the new age architecture is going to be PLC. will be virtualized. It'll sit in a cognition box kind of a strategy so that you have a direct access to the PLC. In fact, if if that happens, which will happen very soon, uh, you'll be able to move the machines directly. You have a controls at the
- 17:07
- Unidentified speaker (speaker_1): PLC level. So that's the one level of transformation or redesign will happen. Second, it's about where you where you transact this conversation, meaning if you deal at an operational level, at a deep operational level, they will struggle on a basics,
- 17:27
- Unidentified speaker (speaker_1): not to say anything bad about them, because simple thing about how do I uh open the router and give you an access to you for my system, or how do I open the protocol of the third third-party party system to you?
- 17:40
- Unidentified speaker (speaker_1): So I think these conversations, the new age conversation, are going to start probably at the CXO level so that they see a different view altogether that requires a different upgrade strategy, infrastructure upgrade strategy. And then they actually see as a long-term benefit, not as buying a tool short-term. short term. So this
- 18:00
- Unidentified speaker (speaker_1): maturity is going to happen through an evolution journey.
- 18:05
- Unidentified speaker (speaker_1): It requires to have a bigger transformation agenda,
- 18:09
- Unidentified speaker (speaker_1): a broader vision on driving the autonomy.
- 18:13
- Unidentified speaker (UNKNOWN): And
- 18:13
- Unidentified speaker (speaker_1): I think Honeywell had been very selective in what we call a lighthouse customer of the first 15-24. And these customers, we start with the CEO or a CXO level, and then we work from top to bottom. Uh, it won't happen in a traditional business overnight.
- 18:30
- Mustafa Okur: Makes sense. So it sounds like there could be some sea change to the whole architecture here. Going down to sort of more specific examples, maybe, could you give us one manufacturing decision maybe or that a human operator makes today that you think an AI agent could make now? Or maybe something that they could make in the next sort of six to twelve months?
- 18:50
- Mustafa Okur: Any specific examples?
- 18:52
- Unidentified speaker (speaker_1): I can give share two tangible, one in a process industry,
- 18:55
- Mustafa Okur: a
- 18:55
- Unidentified speaker (speaker_1): plant environment, the other one in a commercial building environment,
- 18:59
- Unidentified speaker (speaker_1): where we build systems for uh process operator, a classic plant, which we had a press release six months ago from a Middle the East, where it's one of our largest ethylene plants.
- 19:15
- Unidentified speaker (speaker_1): where you have millions of sensors connected. And then being it's a deterministic system, it produces what they call false positive, and you can call them as a nuisance alarm. And the operator spends most of their time in clearing what's false pass what's nuisance alarm and what's the real alarm, and then that's how they they prioritize,
- 19:35
- Unidentified speaker (speaker_1): dissect, and then take decisions. Now we have actually deployed things, what we call plant cognition, which pulls all of them based on the intelligence, prioritizes, dissects it, and then eliminates those noises out and gives the direction to the to the operator.
- 19:55
- Unidentified speaker (speaker_1): So you're eliminating the huge waste and you're talking about somewhere closer to 60 to 70 percent, close to 100,000. false alarms coming up every month.
- 20:05
- Mustafa Okur: Wow.
- 20:06
- Unidentified speaker (speaker_1): In second case, uh, which you're aware of, you know, when we started with our pilot with Nvidia,
- 20:13
- Unidentified speaker (speaker_1): your building management system and security and access control system, security access control system produced close to half 250,000 false alarms every quarter, and then your BMS system would deliver somewhere close to
- 20:26
- Unidentified speaker (UNKNOWN): the same.
- 20:27
- Unidentified speaker (speaker_1): 100,000 false alarms. But then when you really dissect it, you have an ability to really take 85, 95% of false alarm away from your system.
- 20:37
- Mustafa Okur: Now
- 20:38
- Unidentified speaker (speaker_1): the intelligence are ready today. So two things that's happening in place. The intelligence with AI model in place, if you had if you train them to the highest accuracy, it's almost like you're you going to Outlook where you're actually segmenting. between focused email and then the junk email. That's exactly what you're doing today. It's helping in two ways. It's not about
- 20:58
- Unidentified speaker (speaker_1): reduction of process operators, but you're putting process operators to focus on the higher value.
- 21:03
- Mustafa Okur: Yeah. Signal
- 21:04
- Unidentified speaker (UNKNOWN): to noise.
- 21:04
- Unidentified speaker (speaker_1): Yeah. They
- 21:05
- Unidentified speaker (UNKNOWN): are
- 21:05
- Unidentified speaker (speaker_1): now they can spend time on throughput of a building or throughput or a yield of a plant rather than segmenting hundreds and thousands of uh uh news and alarms, segmenting them, analyzing them. Today I think the models have reached a an intelligent quotient where it not only detects
- 21:25
- Unidentified speaker (speaker_1): what's false positive, it has the reasoning, enough data, but we give a power back to the operator to say, tell us whether you agree or not. As long as we keep the pro operator in the loop, I think it's a great evolution.
- 21:39
- Unidentified speaker (speaker_1): into what what I call semi autonomous mode. And it prepares uh the the the space very, very well for the future autonomous move.
- 21:48
- Mustafa Okur: Makes sense. And let's stick with the buildings.
- 21:51
- Unidentified speaker (speaker_1): Yeah.
- 21:51
- Mustafa Okur: Um
- 21:52
- Unidentified speaker (UNKNOWN): I mean
- 21:53
- Mustafa Okur: you did talk about what does autonomous actually add. It definitely improves sort of the signal to noise ratio for human operators.
- 21:59
- Unidentified speaker (UNKNOWN): What
- 22:00
- Mustafa Okur: do you think will a state state-of-the-art of the art building do? differently apart from that if you have a view, you know, five years from now that it cannot do today?
- 22:09
- Unidentified speaker (speaker_1): It's an interesting. We we believe there are going to be five or six agents that are going to be standard offering for any building. Easy, uh direct fit right now is energy uh agent looking at all different energy patterns in a building. and advising and
- 22:29
- Unidentified speaker (speaker_1): directing the operator, either it's a peak load, inefficient energy usage because of a poor uh management of an asset or poor asset maintenance, it provides a lot of insights. So one is straightforward as we go live with all the pilots right now. I think energy takes the first lead. Obviously, the
- 22:48
- Unidentified speaker (speaker_1): expectation is if you're your traditional. building controls plus the first generational forge, if it had delivered a single digit energy efficiency value so far, can the new age agents and autonomous controls can they deliver 2x? That's what we are actually really going through right now.
- 23:08
- Unidentified speaker (speaker_1): Second one, which was interesting, we call operational efficiency agent, which is elementary. eliminating all the nuisance alarms coming from a variety of different controls in a building. That's going to come from a fire control, building management system, security and access control. I think that's a second big
- 23:27
- Unidentified speaker (speaker_1): value system. I think everybody are banking in because having a security operations center, building operations center, manning people to eliminate, eradicate the nuisance alarm, I think that's a the biggest value. Third, which we see that as a comfort uh agent right now, comfort is all about as we are
- 23:47
- Unidentified speaker (speaker_1): in this building right now, uh, how do you ensure that cooling and heating and uh ensuring that there's not enough um you know carbon deposit? So there is a lot of number of things. How do you keep you and me pretty satisfied, happy through the whole day? Fourth, we have not even trial tested yet in the
- 24:07
- Unidentified speaker (speaker_1): early days. As we contextualize every asset in a building, we understand the operating condition of an asset. Can you actually have a cyber protection or decyber around this asset? I think it's a adjacent, we are actually working at this point in time because as I zoom into a building, if I actually detect 10,000 assets. in a
- 24:27
- Unidentified speaker (speaker_1): building, I understand the characteristic of an asset and how the additional values to look at the cyber. I think the fifth one that I actually see is going to come very, very clearly is the security and access control, which we used to have predominantly video camera surveillance-based. Everything is moving into a software-based. software based.
- 24:47
- Unidentified speaker (speaker_1): And how do you really ensure one is you're controlling this. the security, access to the building, access to the zone, access to the floor. So there are four or five things. What's happening is it's fascinating where software, AI, intelligence are taking the lead. The number of use cases, the world has not even seen it,
- 25:07
- Unidentified speaker (speaker_1): which is about to open up. And those things will probably drive. I I actually see two different faces. phases. One, you may actually have a native intelligent building designed from a scratch for all the green field. And then you're going to see a massive upgrade to the brown field projects. How can you really bring in a layer on top of my
- 25:27
- Unidentified speaker (speaker_1): existing infrastructure? Can you actually start to really deliver these new agents that can start to optimize and give us an insight? What can we control? What can we optimize? I think that's the phase face you're going to start. kind of starting see.
- 25:39
- Mustafa Okur: And it sounds like these could be very high sort of ROI investments for the existing buildings because buildings are I mean the numbers change, but somewhere like 30, 40% of global emissions worldwide. And there's a lot of room, especially in a complex commercial building like the one we're sitting in right now, you can have a lot of different zones, right? And there's a
- 25:59
- Mustafa Okur: lot of opportunity probably to learn from the use uses of that particular zone throughout the day. Maybe you dim the lights in a tiny zone or do this or that. That helps a little upgrades that add up for the entire building. Is that the way right way to think about it?
- 26:15
- Unidentified speaker (UNKNOWN): It
- 26:15
- Unidentified speaker (speaker_1): is, Mustafa, I other think we are also learning. We we saw in a couple of buildings where it's designed such a way ten 10 years ago, the heating and cooling
- 26:26
- Unidentified speaker (UNKNOWN): uh
- 26:27
- Unidentified speaker (speaker_1): operates through 24 by 7.
- 26:29
- Mustafa Okur: Yes.
- 26:29
- Unidentified speaker (speaker_1): And we went back and said if your temperature outside is more than 58 degrees Fahrenheit, you can switch off the heating system because you don't need it.
- 26:38
- Unidentified speaker (UNKNOWN): And
- 26:38
- Unidentified speaker (speaker_1): in the cooling system, same way too. Now it was designed such a way to keep the comfort high, but now you have the automatic uh AI agents that are running in. They're coming back and saying, can can't can will you allow me to switch off with the set point optimization because you don't need it. I'll make sure that I have a control in place. So these
- 26:58
- Unidentified speaker (speaker_1): in these the uh scanning of the possibilities are millions. And uh and uh because you're launching more and more software agents, like uh software robots that are running in and giving you an opportunity to uh uh optimize, I think those things are. real at this point in time.
- 27:16
- Mustafa Okur: Great. Um
- 27:17
- Unidentified speaker (UNKNOWN): going
- 27:18
- Mustafa Okur: back to the implementation of all this, at the investor day you talked about connecting environments which can have thousands of assets across different protocols and vendors. How close are we to making that integration effectively sort of plug and play
- 27:34
- Unidentified speaker (UNKNOWN): uh
- 27:34
- Mustafa Okur: rather than a large engineering project that takes days and weeks?
- 27:38
- Unidentified speaker (speaker_1): You know the the one that we showed during Investors Day, it is operational at this point in time, which you're talking about the Voyager headquarters, which we had a recent press release by Nvidia and Honeywell. We talked about openly where Scott and Ray, who runs the real estate and operation for Nvidia,
- 27:58
- Unidentified speaker (speaker_1): talked about the value of forge cognition, the value of autonomy. They speak eloquently on outcomes, which is operational efficiency, energy efficiency, comfort efficiency. We have taken that approach into two new buildings for Nvidia right now, which is going
- 28:17
- Unidentified speaker (speaker_1): to be an interesting new category of building. One is a mini data center, the other one is going to be a complex RD lab. The reason we are trial testing forge cognition on three types of building, Voyager is out, which is a critical infrastructure where engineers and it's a white-collared building, then you have
- 28:37
- Unidentified speaker (speaker_1): a mini data center and we have a mini lab. We want to see how the orchestration of this intelligent brain behind the building and the how the learning is uh propagated between the three buildings. Now you can ask the question, how long did it take to really build it? I think when we showcased during our uh investors' day,
- 28:57
- Unidentified speaker (speaker_1): we had a 90-day challenge to go from
- 29:01
- Unidentified speaker (speaker_1): concept idea to deploying our cognition on top of the existing infrastructure with the layers that we have built, Build, which is our IP protected layer to connect to a multiple asset, we were able to get this on time. And then we took another 90 days to really run the operation and show the real ROI to Ska
- 29:21
- Unidentified speaker (speaker_1): to the operations team. And that's the point in time. We had a press release, uh, got the month of October, I think, to really talk about the value of the autonomous building. So the the answer to your question is we have built a lot of this those infrastructure for the last two years. From now on, it's
- 29:41
- Unidentified speaker (speaker_1): going to be my my expectation is if I get in a building's environment, if I get a new building, within a week's time, we should be able to plug in our platform on top of any existing building, discover all an as all the assets in a building, contextualize them, then spend
- 30:01
- Unidentified speaker (speaker_1): another a month or two. to improve the efficiency and intelligence around those asset and fault models. So within 90 days, we should be able to take a brown field and we should be able to get up and running. In fact, to prove that concept, right now Honeywell is working with 15 different uh
- 30:21
- Unidentified speaker (speaker_1): lighthouse um customers, variety, a a variety going from uh a hospitality. vertical customer, a mini data center vertical customer, stadium in UK and Charlotte. So we're working with variety of them to say if with if this is repeatable, it changes the dimension and game on
- 30:41
- Unidentified speaker (speaker_1): uh building automation pitch moving forward.
- 30:44
- Mustafa Okur: That makes sense. And clearly the ROI was there for NVIDIA so that you know they came back and you now have a mini data center and a lab. Um,
- 30:53
- Unidentified speaker (UNKNOWN): are
- 30:53
- Mustafa Okur: you building a muscle that will sort of help you hasten these developments? You said, you know, first one, two years you built this organization up, so it's already got some capability. You're demonstrating that with the Voyager campus of NVIDIA. Every time you do a project, I assume you're learning as a company. So we do you think that the the deployment will speed up?
- 31:13
- Unidentified speaker (speaker_1): You know, it's gonna be going a to new muscle. muzzle for the industry. I think uh we have today a platform steam. The new t new muscle that needs to be built, you hear a lot about FTE, forward deployed engineers. And this is a cross-functional group who understands the platform, walks into the building,
- 31:33
- Unidentified speaker (speaker_1): you have a system expertise, services expertise who understands the software expertise.
- 31:39
- Unidentified speaker (speaker_1): and uh project implementation expertise. So it's a cross-functional group. And uh we are in uh step one right there. How do you really bring in, build that muzzle a lot? The good news is we have a large project implementation talent within the company, close to almost 4,000 engineers
- 31:59
- Unidentified speaker (speaker_1): supporting buildings and processes process and uh you UOP technology. I think we need to go back and leverage them because they have 20 to 30 years of systems domain knowledge. Then we need to bring in a new age software talent, which I think is easily available because you can bring in a native
- 32:19
- Unidentified speaker (speaker_1): AI expert to fuse them with the domain experts. But the new muzzle for meaning your question is if I'm doing 15 pilots. I need to have uh 15 FDE teams who can live next 90 days within the customer's environment. One instrument,
- 32:38
- Unidentified speaker (speaker_1): operate, show the value before you can walk out and starting to really the the job starts after 90 days because you're trying to sell an outcome. I think the customers are asking you to really monitor. all the key parameters. And then their job, our job is to reinforce two things. Either you reinforce your software
- 32:58
- Unidentified speaker (speaker_1): algorithm to improve the accuracy, or reinforce your process controls within the operation. You're retraining your workforce, you're retraining your operations team. So this is going to be the new muzzle. And uh that's an area that I think we will spend all our time. in the coming 12 months uh to scale
- 33:18
- Unidentified speaker (speaker_1): and uh and learn all together. I think the industry needs it. Today companies like Palantir does it very well. They walk into a customer in six weeks. I think they call a boot camp. Um they can really take their methodology, deploy at scale, and hand over the keys. I think in an industrial environment world, I think we are doing a
- 33:38
- Unidentified speaker (speaker_1): pilot for. us to learn uh how big is that capacity and capability we need to build, and what is our commercial muscle to scale?
- 33:46
- Unidentified speaker (UNKNOWN): So
- 33:46
- Mustafa Okur: yeah, that that makes sense. And there is commercial muscle, which I'm sure Honeywell does have. Uh the other question there would be do you have to retrain the models when you're deploying across different building types? Or I'm imagining more that the models are already there. you're just enriching them with the context of the customer.
- 34:06
- Mustafa Okur: Would that be the right way to think about it? Because training a model would take probably a lot of resources, time,
- 34:11
- Unidentified speaker (UNKNOWN): and so on.
- 34:12
- Unidentified speaker (speaker_1): So there are two steps. One, every assets that we discover, they have their associated fault models. And you may have a different different input inputs that may come from a different asset type. So we may always continue to fine-tune our fault models for every asset type because the accuracy of the
- 34:32
- Unidentified speaker (speaker_1): fault model is so important in managing the asset and the performance of an asset in a building. But the second question that you're asking is if we end up start building a domain specific uh foundation model where all our IPs are well protected with our with our system.
- 34:51
- Unidentified speaker (speaker_1): how often do we need to start fine-tuning based on new customer, new vertical news segment? I think we are again in the early stage today where if we starting to really take our open source, which I said, which is which is going to be a hybrid model fused enough for buildings and fused fuse for a plant operation moving
- 35:11
- Unidentified speaker (speaker_1): forward. I think the approach that we are working much more closely with the model providers like Nvidia or other experts are telling us that maybe at least once a quarter we will be thinking about fine-tuning based on the learnings that we all have. And the the key thing is the muscle that
- 35:31
- Unidentified speaker (speaker_1): we have to build in an industrial context world uh, to ensure that over-the-air upgrade into the industrial environment. So those are the muzzles that we are in the early stage right now.
- 35:42
- Mustafa Okur: But
- 35:42
- Unidentified speaker (speaker_1): I foresee it's not going to be like an Apple every two weeks or every week, like Tesla, you deploy software over the air, but at least once a quarter, your ability to upgrade with patches for OT cybersecurity control model changes or the learnings between the different buildings.
- 36:06
- Unidentified speaker (speaker_1): Those are the things that we are seeing a possibility, Mustafa, which we were starting to experiment starting Q1.
- 36:12
- Mustafa Okur: I'm glad you brought up over-the-air updates, Rush, because um, well, as we're recording today, there is another uh cyber attack happening in Europe in uh retail uh supermarkets, uh grocery stores.
- 36:25
- Unidentified speaker (UNKNOWN): Uh
- 36:25
- Mustafa Okur: so cybersecurity is definitely becoming more and more important. So it's great to hear that, you know, maybe not like an iPhone or uh sort of Linux servers that get an update all the time, but every quarter it does make sense to have these buildings.
- 36:39
- Unidentified speaker (speaker_1): And I want to assure you and the listeners, the moment I say over there upgrade, I'm sure the question will pop up to say, okay, how do we how do we know your device is well protected? Uh we partnered with Google very early and we decided that we'll go with Air Gapped approach. So GDC from
- 36:59
- Unidentified speaker (speaker_1): Google is the one who will be as part of our Bill of Materials spec in Forge Cognition. So because Air Gapped is a mandated one when you want to deal with federal defense and any of the government agencies in the future. I see that across the world. So uh we have gone ahead upstream to have GDC,
- 37:19
- Unidentified speaker (speaker_1): uh that's part of our Google partnership and strategy.
- 37:22
- Mustafa Okur: That makes perfect sense. And that answers my next question as in how would you use defenses maybe using AI uh in a world where we're seeing more and more cyberattacks, but air gapping makes probably the case
- 37:34
- Unidentified speaker (UNKNOWN): of One the clearest
- 37:35
- Mustafa Okur: solution. solutions.
- 37:36
- Unidentified speaker (UNKNOWN): One
- 37:36
- Unidentified speaker (speaker_1): is air gap, so that I want to assure our customers. And I think I I I thank Google from a day one when we had conceptualized the idea and uh it was a co-creation with them. The second part of it, the principle, is as long as we have the deterministic controls
- 37:56
- Unidentified speaker (speaker_1): architecture, and then I'm bringing in a probabilistic where I'm not giving a power and a control to these agents to drive. actions on its own, but still keeping the operator in the loop or
- 38:08
- Unidentified speaker (UNKNOWN): a
- 38:08
- Unidentified speaker (speaker_1): human in the loop. I think that's our commitment to the industry, two things. We are not here to eliminate the work. We are here to really improve the efficiency of the work.
- 38:17
- Mustafa Okur: Yeah.
- 38:18
- Unidentified speaker (speaker_1): And we believe in a mission critical and safety critical world, you need to have the process operator always guiding and directing indecisions. In the past, if they are able to do just 10 actions per day after eliminating all their false positive, today they should be able to process maybe 100 or 200. And uh and that's the efficiency
- 38:37
- Unidentified speaker (speaker_1): lever that we are foreseeing at this point in time.
- 38:40
- Mustafa Okur: Makes sense. Earlier you brought up ontology and you did mention that you know during your phase one of basically adopting AI solutions and autonomy,
- 38:50
- Unidentified speaker (UNKNOWN): you
- 38:50
- Mustafa Okur: went through the this phase of codifying Honeywell's hundred plus years of knowledge internally. Could you tell us you know what that involves in in real life? How does that how is that executed? And is there an opportunity here to do the same for your customers, right? Because they have maybe years or decades of internal knowledge that they're looking
- 39:10
- Mustafa Okur: to codify. Are you are you the right company as an industrial company? industrial to industrial, to do that? Or can that be done by, I don't know, as you said, a Palantir or maybe a Google?
- 39:22
- Unidentified speaker (speaker_1): I think that's an area that you'll go with us or OEMs, because we when I say domain, you absolutely understand you're building uh uh building operation, security and access control, fire, wherever we own our core product and assets. So phase one, we have gone ahead. There are close to
- 39:42
- Unidentified speaker (speaker_1): 130 different asset types in a world of all possibilities and probability in a BMS plus security and access control and fire. There are close to 130 asset types, which we have gone ahead and codified and contextualized, and we have deployed at scale in all our
- 40:01
- Unidentified speaker (speaker_1): pilots today. And I think with the right level of fault models for these asset types, uh, the way we measure is its accuracy in predicting the asset performance, predictive maintenance, and uh and the intelligence questions are measured based on the fault models. Then we actually realized
- 40:21
- Unidentified speaker (speaker_1): uh that for us to move much more closer into certain vertical like data center. you got to have the electrical tax contextualization is critical. As I as you probably heard, we had a partnership with LS Electric a year ago. So we are working with them with our partnership. We
- 40:41
- Unidentified speaker (speaker_1): took their additional 40 to 48 asset category types. We are contextualizing because now as we go to market together with them, so that our ontology layer will cover not only on a building operation side, but also bring in electrical product side. And then now you actually say if you need to go much more
- 41:01
- Unidentified speaker (speaker_1): further into a into a data center, then there are liquid cooling slash power generators, slash UPS, slash IT racks. I think we want to really partner with a few OEMs who want to partner with us. And if we can actually really finish, I think there are probably additional hundred more asset types. So what's
- 41:20
- Unidentified speaker (speaker_1): happening is, Mustafa, as we move from our core domain and we are moving into a new vertical, we're starting to realize it requires a consortium of partners who may bring in their OEM products, willing to share their information and data and I think if there is one layer
- 41:40
- Unidentified speaker (speaker_1): we're going to build moving forward, which affords cognition, ontology is our intelligence layer. And that intelligence layer is going to come with our partners, our core sponsors in our ecosystem. I think, I think we are making progress. We are making progress and this is going to be an evolution journey. And what
- 42:00
- Unidentified speaker (speaker_1): we're really doing is what's what's what what's brilliant about this is you are actually retraining the intelligence in a physical world that existed in people's minds and a probably discrete set of intelligence light across the board. We are codifying them back. What happened to the
- 42:20
- Unidentified speaker (speaker_1): GPT world of the first generation, they took internet data and trained the model. But here we are taking a hardcore physical world data, we're training it. I think I'm actually starting to see a progress. And I think for me,
- 42:34
- Unidentified speaker (speaker_1): recreating this for the process industry, where it's going to be in industrial assets. We just announced, if you have seen it, the press release two weeks ago with Dangote,
- 42:45
- Unidentified speaker (speaker_1): the uh process plant in Nigeria. It gives us a unique opportunity. We're going to gonna take take Honeywell. Honeywell, Nvidia, and Dell going to do this with Dangote's leadership team. We're gonna deploy forge cognition right there. What's the beauty of that implementation would be industrial assets and also UOP core
- 43:04
- Unidentified speaker (speaker_1): process units.
- 43:06
- Unidentified speaker (speaker_1): The complexity in process units is you need to take your chemistry, your catalyst information and formulae. you've got to codify it. So it's a first generational problem we're going to solve. If you do so, I think the kind of a problem we may solve in an industrial world will be probably a multiple X
- 43:26
- Unidentified speaker (speaker_1): in ROI to add a commercial building today.
- 43:29
- Mustafa Okur: That makes perfect sense. It reminds me of the uh acquisition Google did of uh is it the spirit uh booking management system? Basically. to get the data to add to their ont ontology of the models. So
- 43:42
- Unidentified speaker (UNKNOWN): what
- 43:42
- Mustafa Okur: you're describing there makes perfect sense to me.
- 43:45
- Unidentified speaker (speaker_2): That's
- 43:45
- Unidentified speaker (UNKNOWN): right.
- 43:46
- Mustafa Okur: Um I'm glad you brought up data centers, Suresh, because you know we can't really chat with anyone without uh data centers these days days. Um could you help us understand where Honeywell participates in a data center today? From it could be from design all the way to like fire and safety.
- 44:00
- Unidentified speaker (speaker_1): Today, look I think we have a pretty solid business backlog. Uh we're growing that vertical so nicely, cross-selling multi-portfolio. We have sensor business selling um sensors for liquid cooling to building uh from building automation business,
- 44:20
- Unidentified speaker (speaker_1): selling fire security, access control, building management system. So we have a wide variety going for across the portfolio. But we also understand the future. You got to start to look at the new AI reference architecture, the specter on the architecture. How do you really make sure that
- 44:40
- Unidentified speaker (speaker_1): those are designed for the specification? So we started our initial conversation with Nvidia, and we understood between Vera Rubin and Ultra and Fenman, which is going to come up in 2028, how should we be thinking about the future state of reference design? How do you redesign your controls architecture? How do you
- 44:59
- Unidentified speaker (speaker_1): redesign your product that also
- 45:03
- Unidentified speaker (speaker_1): addresses the need? How should you be re-realizing some of the core products that need closer to a data center vertical? So we are in the early discussion at this point in time that may also lead into, I'm sure you, you know, you may have. a question on recently. Nvidia also talked about modularizing their architecture design.
- 45:26
- Unidentified speaker (speaker_1): I think it's it's it's brilliant, revolutionary. And we ended up sitting and talking about that reference design. How should we rethink about our battery energy storage and our building control? controls architecture? Can they coexist into the reference design? So the the focus for us is so
- 45:46
- Unidentified speaker (speaker_1): far, we have sold multi-portfolio into the data center vertical, but now we also believe the future is going to be reference design spec, so that you're designing for the spec and you work with the vertical. We do have one big customer. Equinix. We are working much more diligently with
- 46:05
- Unidentified speaker (speaker_1): them and which is our strategic customer. But uh we realized that you got to starting to understand the spec and we knew that the future is all about that. And that that that allows it's not only for specing and you know you be part of it, but the future is all about standardization. Multi Multi-vendors vendor should abide by those standard
- 46:25
- Unidentified speaker (speaker_1): protocols and standards standard subsystem interfaces that will help because data centers of the world has to deliver uh 99.9999%, which I talked about.
- 46:36
- Mustafa Okur: Yeah.
- 46:36
- Unidentified speaker (speaker_1): And that requires multi-vendors has to abide by the spec and standards. So
- 46:40
- Unidentified speaker (UNKNOWN): we
- 46:40
- Unidentified speaker (speaker_1): are in the early days at this juncture. You're going to see a lot more in this particular space.
- 46:44
- Mustafa Okur: Yeah, and I think um,
- 46:46
- Unidentified speaker (UNKNOWN): you
- 46:46
- Mustafa Okur: know, being an official partner and ref in the being in the reference design, it does help. and there are companies who do that, but ultimately I think the key thing is being able to meet the spec because the the power consumption is reaching such levels, especially with Feynman, and the thermal envelope is so steep,
- 47:06
- Unidentified speaker (UNKNOWN): it's
- 47:06
- Mustafa Okur: because these things will run hot.
- 47:08
- Unidentified speaker (speaker_1): That's right.
- 47:09
- Mustafa Okur: And the cost of having a fire in a data center will be so enormous. that ultimately you need a dependable partner who can meet the specs. That's the key thing here. The other thing people don't seem to understand, and please correct me if this is wrong or or partially true, you can have reference
- 47:29
- Mustafa Okur: specs from NVIDIA for these modular designs, the DSX pods they call
- 47:33
- Unidentified speaker (UNKNOWN): them,
- 47:34
- Mustafa Okur: but ultimately the hyperscalers do whatever they want to do, right? So um you
- 47:38
- Unidentified speaker (UNKNOWN): can
- 47:38
- Mustafa Okur: have an Amazon uh data center which will look very different from a meta center uh data center. So it's great to hear that you have a seat at the table and thinking about this going forward. Is the primary challenge there, would you say, the physics of it, like the the megawatts that are being pushed? Is it the variety of the vendors
- 47:58
- Mustafa Okur: who are playing in this field? Maybe you foresee a consolidation. What are the biggest sort of challenges?
- 48:05
- Unidentified speaker (speaker_1): I think for me it's going to be the uh the second one. I I'm actually happier that this uh consolidation through this DSX partnership that NBA is driving, because I'm assuming that in the past every vendor were trying to build for their own specification for their own industry, and we were trying to probably cross crossle it to
- 48:25
- Unidentified speaker (speaker_1): this vertical. But but somebody taking the lead, I'm sure AMD is going to define their own reference design. Nvidia is doing it already. I'm sure if if Google wants to really start into push TPUs out, they need to have the reference design for the hyperscalers. There are two types you have hyperscalers and then you have neo cloud providers and then you
- 48:45
- Unidentified speaker (speaker_1): also have probably a mid-scale mid scale cloud providers. I think my view is across the board, people need to have for two reasons. One, resiliency of these data centers. Second, you're right, these are these are going to be burning a lot more energy and uh somebody if you don't
- 49:05
- Unidentified speaker (speaker_1): if they don't really manage the thermal uh cycle and uh you know cooling cycle, I think it's going to have a uh tremendous ROI problems in the future. So I I genuinely believe, I think it's just a maturity of the market. It's evolving. The fact is, I'm telling you right now, I would have said it six months ago, is
- 49:25
- Unidentified speaker (speaker_1): we are at the seat talking about the next architecture. We are really looking at that much more closely, and is actually forcing us to go back and look at our own roadmaps. What should we spec for that design? What changes we need to tweak? Third, are there any other portfolio we may also need to augment? And the
- 49:44
- Unidentified speaker (speaker_1): fourth one is: should we actually have some partnership much more closely with others? Because not everybody can solve every other equation in every data center. There are eight different building blocks. We come across around software and controls layer, but then you have to partner with another three or four different building blocks in a data center. to be a uh to
- 50:04
- Unidentified speaker (speaker_1): to be a a robust partner to deliver a value.
- 50:07
- Mustafa Okur: That makes sense. And and I
- 50:09
- Unidentified speaker (UNKNOWN): you know,
- 50:10
- Mustafa Okur: you of course you can't disclose details, but does that mean you're talking to some of your industrial peers as well? And here I'm not thinking about a conglomerate peer, but a a specialist peer, let's call it a vertive, right? So they do
- 50:22
- Unidentified speaker (UNKNOWN): a
- 50:22
- Mustafa Okur: lot of things inside the building, but not necessarily the controls of the building.
- 50:26
- Unidentified speaker (speaker_1): Yes. Meaning you'll hear more about in the coming months. But those are the consolidation we need because there are strength that we bring to the table. There are strength that others bring to the table. If you can really, really blend them together, Mustafa, it's going to be a huge value. I uh and uh I I see the potential that way.
- 50:45
- Mustafa Okur: That makes perfect sense.
- 50:47
- Unidentified speaker (UNKNOWN): Well,
- 50:47
- Mustafa Okur: look, it's it's rare to find company executives who will dive. into technical details like we did today, Sir. So
- 50:53
- Unidentified speaker (UNKNOWN): uh
- 50:54
- Mustafa Okur: it's a breath of uh fresh air. It's clear that a lot of thought has gone in into where you have chosen to participate, where you're thinking about partnering, just as we talked about now for the company. And uh the outcome is more sort of productive and secure solutions for your customers who are global but actually skewed to the US. So
- 51:14
- Mustafa Okur: it's great for uh everyone involved, including the country. Thank you very much for the time today, for taking the time today to sit down with us and clarify these new sea changes that we're seeing across industrials. And thank you to our listeners.
- 51:30
- Unidentified speaker (UNKNOWN): Uh
- 51:30
- Mustafa Okur: this is Mustafa Oakur. Uh thank you for tuning in to another episode of Tech Disruptors.
- 51:35
- Unidentified speaker (speaker_1): Thank you so much, Mustafa.
More from Tech Disruptors
Figma on Why Design Gains Value in the AI Era
Episode discussion: Kris Rasmussen on Figma's AI Strategy: Design Context as a Moat and Hybrid Pricing Models
Speakers’ claims: Kris Rasmussen, Chief Architect at Figma, discusses how AI agents serve as interface glue rather than replacements for traditional UIs, emphasizing that design systems provide critical context for better AI outputs. He argues that as code becomes commoditized, value shifts to design, with Figma leveraging proprietary user data to fine-tune open-weight models for specific workflows. Rasmussen highlights bottlenecks in AI adoption due to security and backend expertise gaps, noting that human curation remains essential. He also outlines Figma's infrastructure strategy, which avoids building foundation models from scratch, and predicts a shift toward hybrid seat-based and consumption-based pricing models.
IBM CEO on Enterprise AI and Software Strategy
Episode discussion: IBM CEO Krishna outlines enterprise AI adoption stages, open-weight model economics, and software consolidation trends.
Speakers’ claims: IBM CEO Arvind Krishna discusses the state of enterprise AI, noting that only 10-20% of companies have moved from experimentation to scaled deployment. He identifies coding, customer experience, and enterprise operations as key use cases, predicting that tech spend will double as a percentage of revenue over the next decade. Krishna argues that data accessibility matters more than cleanliness for AI agents and highlights IBM's internal success with AI automating HR and procurement processes. He forecasts that open-weight models will capture 80% of enterprise token usage within three years due to cost and control advantages, drawing parallels to the adoption of Linux.