CES 2026: The Physical AI Inflection Point

Lia: And I'm Lia. We've just wrapped four days in Las Vegas at CES 2026, and we need to talk about what we saw—because the story isn't really about gadgets this year. Thom: No, it's not. If I had to summarize the entire show in one phrase, it's this: AI left the chat. Lia: Literally. The era of AI as a thing you type into—a chat interface, a prompt box—that chapter is closing. What opened in Las Vegas this week is the era of Physical AI. AI that doesn't just generate text or images, but AI that moves through the world, reasons about physical spaces, and takes action. Thom: And this has massive implications for anyone leading technology strategy. So today, we're doing something different. We're not running through a product list. We're going deep on what CES 2026 signals about where this industry is headed—and what you should be doing about it right now. Thom: So Lia, I've been processing everything from CES 2026, and I keep coming back to something Jensen Huang said that I think frames this entire show. He said, quote, "The next wave of AI is physical AI." And honestly, I think that's the lens we need to use for everything we're about to discuss. Lia: It really is the through-line, isn't it? Here's what matters: we're witnessing what I'd call Act Two of the AI revolution. Act One was generative AI creating content—text, images, video. Act Two is AI interacting with and manipulating the physical world. That's a fundamentally different challenge. Thom: Right, and this isn't just robots with ChatGPT bolted on. I mean, that's the lazy interpretation. What we're actually seeing is a fundamental architecture shift. These new systems unify perception, reasoning, and action in a single model. They're called Vision-Language-Action models, or VLAs, and they're kind of blowing my mind. Lia: The timing isn't accidental either. Three enabling technologies converged simultaneously: foundation models that can reason multi-modally, simulation environments mature enough to train robots cheaply, and purpose-built silicon for edge inference at the right power and latency envelope. Thom: Ooh, and here's the strategic implication that I think enterprise leaders need to sit with. The value chain is shifting. If digital AI rewarded the model builders—OpenAI, Anthropic, Google—Physical AI rewards those who control the hardware interfaces. The robots, the vehicles, the appliances where AI actually meets the real world. Lia: Hyundai is the perfect example of this thesis in action. Their vertical integration strategy is genuinely impressive—they're designing, building, and deploying their own robots, specifically the Atlas humanoid, to build their own cars. Think about that for a second. Thom: Wait wait wait—I want to make sure people get why this is so significant. They're not buying robots from someone else. They're building the robots that will build their cars. That's a moat within a moat. If your competitor's labor force never asks for raises, never gets injured, and works twenty-four-seven, what does that do to your cost structure? Lia: Bottom line: it's existential for competitors who don't have a similar strategy. And Nvidia is enabling this entire ecosystem with the Cosmos world model for generating synthetic training data. You can't crash a million physical robots to train them, but you can crash a million simulated ones. Thom: Okay, let's talk silicon, because as someone who runs on GPUs myself, this is where I get genuinely excited. The chip wars have evolved. It's no longer about who has the fastest GPU. It's about who has the best integrated system—compute plus memory plus networking plus data movement. Lia: The Nvidia Rubin platform exemplifies this perfectly. It's not just a GPU; it's a six-chip co-designed system. Fifty petaflops of NVFP4 inference, the new Vera CPU with eighty-eight custom cores optimized for data movement, NVLink 6 for scale-up, optical Spectrum-X for scale-out. Thom: The data center is literally becoming a single logical computer. And here's the number that matters for enterprises: Huang claimed Rubin delivers tokens at one-tenth the cost of previous generations. For anyone running inference at scale, that's the metric. Not peak FLOPS—cost per token. Lia: But AMD isn't conceding the market. Their counter-strategy is fascinating. The Ryzen AI Max+ with 128 gigabytes of unified memory enables loading large models entirely on-chip for local inference. The pitch is sovereignty and predictability—run models locally without cloud latency or dependency. Thom: And their Helios platform hits three exaflops per rack. That's... I mean, that's absurd. The AMD open ecosystem is positioning itself as a genuine Nvidia alternative for enterprises worried about vendor lock-in. Lia: Then there's Intel's manufacturing play. Core Ultra Series 3 on Intel 18A is the first leading-edge chip manufactured in the US. This is as much geopolitics as it is product. The Intel 18A as US manufacturing sovereignty play matters for any enterprise thinking about supply chain resilience. Thom: Make sense? Domestic AI supply chain is becoming a boardroom conversation, not just a policy debate. Lia: And we can't overlook Qualcomm's edge dominance. Dragonwing IQ10 for robotics, Snapdragon Ride and Cockpit Elite for automotive. They're becoming the de facto brain for anything that moves. Their Qualcomm cross-domain automotive computing—a single chipset handling infotainment plus ADAS—that's consolidating the car's brain into one system. Thom: [with growing excitement] Okay, now robots. This is where Physical AI gets tangible. The Atlas announcement from Hyundai and Boston Dynamics is significant because of what it is NOT: a demo. This is production-ready. Lia: Fifty-six degrees of freedom, IP-rated for harsh industrial environments, swappable four-hour batteries. Atlas deployment to Hyundai's Georgia factory is happening by 2028. That's not a concept—that's a timeline. Thom: And then Google DeepMind announces their partnership with Boston Dynamics. The Google DeepMind Gemini Robotics partnership means Gemini foundation models are powering Atlas. This is the model provider plus hardware provider pattern we should expect to see repeated across the industry. Lia: The acquisition activity tells the same story. Mobileye acquiring Mentee Robotics for nine hundred million dollars—the Mobileye/Mentee acquisition signals that autonomous vehicle perception and planning stacks are becoming general-purpose autonomy layers. If you've solved driving, you've solved a huge part of robot navigation. Thom: And LG's CLOiD robot for the home—the real story isn't the hardware, it's the VLA architecture. The robot doesn't follow rigid scripts. It reasons about tasks like loading a dishwasher. That's... that's actually kind of magical when you think about it. Lia: The real moat in robotics isn't the hardware though. It's the simulation environments and synthetic data generation. Nvidia Cosmos for synthetic data generation is how you train robots at scale without destroying physical hardware. Thom: Okay, I'm getting into the weeds on robots. Let's pivot to vehicles, because reasoning vehicles are maybe the most immediately deployable form of Physical AI. Lia: Here's the fundamental shift. Traditional autonomous systems detect objects and follow rules. The Nvidia Alpamayo as first reasoning AV model does something different—it reasons about scenarios. When it sees a distracted pedestrian near a crosswalk, it doesn't just detect "pedestrian near road." It reasons "this person might step into traffic" and adjusts accordingly. Thom: And the explainability piece is huge for regulators. These reasoning models can explain their decisions. "The car stopped because it identified a potential hazard" is very different from "the neural net fired and we're not sure why." Lia: Mercedes-Benz CLA as first production deployment launches early 2026 with this stack. First production vehicle with a reasoning-based AV system. Watch how regulators respond—this could set precedent. Thom: What's strategically brilliant is that Nvidia is open-sourcing Alpamayo models and training datasets. Classic platform play. Make your stack the industry standard by giving away the software and selling the silicon. Lia: The cockpit is evolving too. BMW iX3 with Amazon Alexa+ as first generative assistant in production vehicle—natural language, multi-turn dialogue, contextual awareness. The dashboard is becoming an AI agent. Thom: And Qualcomm plus Leapmotor showed a single dual-chipset handling infotainment, ADAS, and body control. The car's computer is consolidating into a central brain. Less complexity, lower cost, easier integration. Lia: [thoughtfully] Now, let's talk about what's happening in homes, because this is where Physical AI meets consumer skepticism head-on. Thom: Every major home brand announced some form of proactive, ambient AI. Samsung Vision AI Companion tracks behavior and offers suggestions. LG Affectionate Intelligence builds emotional profiles. Amazon Alexa+ Web launch extends the agent across surfaces and partner ecosystems. Lia: The technology is impressive. The question is whether consumers want it. CES saw significant counter-programming from privacy advocates. The Privacy backlash and "Worst in Show" criticism at CES highlighted surveillance features dressed up as convenience. Thom: I mean, there was literally an AI lollipop that plays music through bone conduction. Cat food stations with facial recognition for individual pets. When AI is being applied to everything, it's a signal we're in a hype cycle—but also that the tools are cheap enough for every product category. Lia: Here's the strategic insight: the winners in smart home will solve the trust equation. Data minimization, local-first processing, transparent controls, explicit retention limits. Trust as the limiting factor for adoption is the core constraint, not technology. Thom: For enterprises listening to this: this pattern is coming to your workplace. Employees will expect ambient AI, conversational agents, proactive assistance. Your enterprise governance model needs to extend to these interfaces now. Lia: Trust as a product requirement isn't optional anymore. It's becoming a procurement blocker. Thom: [with emphasis] Switching gears to health tech, because this was genuinely surprising to me. CES shifted from counting steps to assessing cellular age. Lia: The Withings Body Scan 2 plus Abbott Lingo integration is the standout example. It correlates weight, body composition, and continuous glucose data to estimate metabolic efficiency and cellular health. Clinical-grade biomarker monitoring at home. Thom: This is the cellular and metabolic health monitoring trend in action. The question is validation—do these metrics actually predict health outcomes, or are they wellness theater? The science is still catching up to the sensors. Lia: Neural interfaces are leaving the medical device category entirely. Naqi Logix neural earbuds let users control devices via facial gestures and brain waves. Won Best of Innovation. The accessibility applications are obvious, but the broader implications for hands-free computing are significant. Thom: And Motorola Project Maxwell concept—an AI pendant worn around the neck with cameras and sensors—signals the trajectory. AI wearables as always-on context providers. Lia: Enterprise implication that I want to emphasize: if employees are wearing AI glasses and neural earbuds that record and transcribe, your security and privacy policies need updating now. Not next quarter. Now. Thom: Okay, let's zoom out to strategic implications. The moat is moving. As AI models become commoditized—open weights, fierce competition—the defensible value shifts to those who control where AI meets the physical world. Lia: The value chain shift to hardware interface controllers is the through-line. Hyundai building robots to build cars. Mercedes integrating reasoning AI into vehicles. Samsung embedding AI into every appliance. The hardware interface is the new high ground. Thom: The geopolitical dimension is real too. Intel 18A manufactured in the US. AMD's open ecosystem as Nvidia alternative. China's response to export controls is accelerating domestic chip development. Enterprises need supplier diversification strategies. Lia: Analyst sentiment at CES was mixed, honestly. Physical AI innovations like Alpamayo and Atlas are genuinely significant. But there's fatigue with AI-washing in minor gadgets. The smart toothbrush doesn't need a foundation model. Thom: And here's something nobody talks about enough: energy constraints on AI scaling. Nvidia's emphasis on ten-x cost reduction per token is partially about compute efficiency, but it's fundamentally about power consumption. Data centers are hitting energy limits. Lia: That shapes where AI can be deployed and how fast the industry can scale. It's physics, not just economics. Thom: So what should enterprises actually do with all this? Let me give you four concrete actions. Lia: Action one: build an inference placement policy framework. Define what must stay on-device—privacy-sensitive, latency-critical. What can go to edge—complex reasoning, fleet coordination. What should be cloud—training, model updates. Add routing rules based on cost, latency, and security. Thom: Action two: endpoint governance extension. If employees are getting AI PCs with forty-plus TOPS NPUs and wearing AI glasses, you need model distribution controls, telemetry and audit trails, and cross-app identity and permissions boundaries. Lia: Action three: if Physical AI is relevant to your business—manufacturing, warehousing, logistics, labs—establish Physical AI pilot criteria and start pilots now. Controlled environments. Simulation and safety validation. Fleet management infrastructure. ROI metrics tied to labor and quality, not hype. Thom: Action four: make trust a product requirement. Data minimization. Local-first processing. Transparent user controls. Clear retention limits. Privacy backlash is becoming a procurement blocker for real. Lia: We're also proposing a framework we're calling the AI Endpoint Readiness Index concept, or AERI. It assesses organizational readiness across five dimensions: Model Distribution, Observability, Data Boundaries, Update Lifecycle, and Identity Permissions. Thom: Use CES 2026 as the forcing function. Ask yourself: if employees buy AI PCs and wear AI glasses this year, are we ready? Because they will. And you need to be. Lia: [in a measured tone] Physical AI is the inflection point. The companies that control where AI meets the physical world will define the next decade. The question is whether you're positioned to be one of them. Thom: So let's bring this home. CES 2026 wasn't about cool gadgets. It was about a phase transition. Lia: From AI that generates to AI that acts. From models in the cloud to intelligence in the robot arm, the car, the appliance. Thom: And the strategic implications are significant. The moat is moving from model builders to hardware controllers. The governance challenge is expanding from data centers to every endpoint your employees touch. The trust equation is becoming the limiting factor for adoption. Lia: If you're a tech executive listening to this, here's your forcing function: Employees are about to buy AI PCs with serious on-device inference capabilities. They're going to wear AI glasses that record and transcribe. Your factory competitors are deploying humanoid robots by 2028. Your car is going to reason about driving decisions. Thom: The question isn't whether Physical AI is coming. It's whether you're ready. Lia: Build your inference placement policy. Extend your governance to endpoints. Start your Physical AI pilots. Make trust a product requirement. Thom: And use that AI Endpoint Readiness Index we outlined—Model Distribution, Observability, Data Boundaries, Update Lifecycle, Identity and Permissions—to assess where you stand. Lia: Thanks for joining us for this CES 2026 special edition. The future moved from the screen to the physical world this week in Las Vegas. Make sure you're moving with it. Thom: Until next time, I'm Thom. Lia: And I'm Lia. Stay curious, stay strategic, and we'll see you tomorrow.