HOST AOK, so we keep coming back to Nvidia. And honestly, that feels right this week. If you wanted one entity that sits at the center of AI hardware, AI software, AI models, and now even AI robots, it’s Nvidia.
HOST BYeah. And what’s interesting is that the conversation isn’t just “Nvidia beat everybody.” It’s more like Nvidia keeps expanding the definition of what counts as Nvidia.
HOST AExactly. And the reason to do a deep dive now is that the last three days were weirdly revealing. Not one giant singular announcement — more like a pile-up of clues about where the company thinks the next phase of AI is going.
HOST BAnd maybe where the bottlenecks are. Because if you look at the news flow, it’s chips, it’s networking, it’s models, it’s PCs, it’s humanoid robots, and even optics. That’s not random.
HOST ANo. It feels like a company trying to own the whole stack, but in a very Nvidia way — not by building everything itself, but by making itself the default substrate underneath everyone else.
HOST B...which is very different from just being a chip vendor. It’s almost like they’re selling the road system, not the cars.
HOST ALet’s start at the beginning, because the backstory matters here. Nvidia was founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem. And the original story is not “we are the AI company.” It’s graphics.
HOST BRight, GPUs. Gaming first, then visualization, then parallel compute. And that evolution matters because AI turned out to be a workload that loved the same thing games loved: massive parallelism.
HOST AAnd that’s the first big historical beat. Nvidia didn’t invent the internet age, and it didn’t invent deep learning. But it arrived with a hardware architecture that was accidentally perfect for the era after deep learning took off.
HOST BUhm, let me think about that for a second... I think the key is that Nvidia kept making the GPU more programmable over time, so when researchers and then companies needed accelerated compute, there was already a platform there.
HOST AYes. And CUDA is the other huge backstory piece. The software layer turned Nvidia from a component supplier into a developer ecosystem. That’s the moat people mean when they say Nvidia is not just chips.
HOST BAnd once the software habit forms, switching gets painful. Because you’re not just replacing silicon; you’re replacing libraries, tooling, debugging workflows, and all the little assumptions baked into production systems.
HOST AThen the timeline accelerates. During the modern AI boom, Nvidia becomes the obvious pick for training large models. First, that was the H100 era in people’s minds. Then Blackwell becomes the next anchor. And now Rubin is on the horizon.
HOST BAnd you can see the company getting more explicit about systems, not just chips. Rack-scale products, networking, full data center design. It’s almost as if the unit of sale is becoming the AI factory.
HOST AThat phrase keeps coming up, and it’s not accidental. Nvidia wants people to think in factories because factories imply throughput, standardization, and industrial scale — not just one-off compute purchases.
HOST BOK, so that gets us to the recent news. The big thing over the last three days is that Nvidia didn’t just do one product launch. It sort of sprayed across the whole map.
HOST AOn June 1, at Computex 2026, they announced a Windows SoC for AI PCs. That’s a pretty clear move into on-device inference and into territory where Qualcomm and Intel are also trying to define the category.
HOST BAnd on the same day, they released Nemotron 3 Ultra, a 550B open-weight model. That’s a huge signal because it says Nvidia does not want to be only the picks-and-shovels company anymore.
HOST ARight. They want to be a model company too, or at least a model-enablement company that can say, “Look, we run the stack end to end.”
HOST BThen there was the humanoid robot reference design with Unitree and Sharpa. That was H2+, and it feels like Nvidia trying to standardize physical AI development workflows the way it standardized training workflows.
HOST AAnd then there’s the optics story from today: Ayar Labs joining the NVLink Fusion ecosystem for co-packaged optics. That one is nerdy, but it may be one of the most strategically important pieces in the bunch.
HOST BBecause it hints at where bottlenecks are moving. If compute is getting denser, then moving data around the system starts to matter as much as raw FLOPS.
HOST AExactly. And meanwhile, SemiAnalysis basically dunked on Jensen Huang’s Computex keynote, saying it was “F tier” because there was no major AI datacenter news. That tells you what the market wanted: not more aspiration, but more infrastructure.
HOST BAnd they also flagged a delayed NVIDIA ARM chip with broken video output. Which is funny, but also important: even for Nvidia, the non-datacenter bets still have execution risk.
HOST AThat’s the pattern, though. Nvidia is still overwhelmingly associated with datacenter AI, but the recent announcements show a deliberate attempt to widen the base. PCs, robots, models, optics, CPUs.
HOST BLet’s slow-walk the business logic, because this is where people can get lost. Nvidia makes money when someone buys compute. But the deeper business is that it tries to sell the entire environment in which compute becomes easy to deploy.
HOST ASo for a non-technical listener: imagine you’re not buying a hammer, you’re buying the hammer, the nails, the workshop, the power grid, and the instruction manual.
HOST BYeah. And the reason that matters is margin, lock-in, and speed. If Nvidia can make an AI system easier to stand up than a competitor’s, then the customer is paying for reduced friction, not just silicon.
HOST AAnd the recent restructuring of reporting is a clue too. They split data center revenue into Hyperscale and ACIE segments. That suggests the company is getting more granular about who the customers are and how the money flows.
HOST BThat’s a subtle but important move. It means Nvidia isn’t just reporting a giant blob called data center and hoping everyone is impressed. It’s trying to map the market structure more precisely.
HOST AThen there’s the Q1 FY2027 revenue number: $81.6B, with networking up 199% to $14.8B. That networking figure is a giant tell.
HOST BBecause networking is not an accessory anymore. It’s core to whether the cluster works. If your GPUs are fast but your interconnect is bad, the whole system underperforms.
HOST AAnd that’s why the Ayar Labs news matters. Co-packaged optics is basically about reducing the pain of moving data across ever-larger AI systems. For a non-technical listener: think of it as trying to keep the roads inside the data center from becoming traffic jams.
HOST BOr put differently: if GPUs are the muscle, networking is the nervous system. And Nvidia keeps buying, partnering, or standardizing around both.
HOST ALet’s talk about Vera Rubin, because that’s the next big hardware storyline. The Vera Rubin NVL72 platform was announced on May 20, and the company is claiming 10x lower cost per token than Blackwell for agentic AI.
HOST BThat claim is obviously directional, but the framing matters. Cost per token is the metric the market actually cares about now, not just peak throughput.
HOST AAnd then on May 23, the Vera Rubin rack was priced at $7.8 million. Which sounds absurd until you remember that Nvidia wants customers thinking in terms of total system economics.
HOST BExactly. The headline price is huge, but the pitch is: if this rack produces enough tokens, enough agents, enough workload, the economics justify it.
HOST AThen on May 27, Phoronix published Vera CPU benchmarks against Intel and AMD server CPUs. That matters because Nvidia is increasingly not just a GPU company, but a system company with its own CPU story.
HOST BAnd this is where I think people underestimate them. If you own GPU, CPU, networking, software, and the rack design, you can optimize the whole pipeline together.
HOST A...which also means if something breaks, the whole stack is your problem. And we saw a version of that with the confidential computing news on May 30.
HOST BRight. Blackwell confidential computing disables NVLink multicast, and that caused a 61% regression on SGLang Qwen3.5 397B. That’s the kind of detail that sounds tiny but is actually huge.
HOST ABecause it shows the tradeoff between security features and performance. For a non-technical listener: imagine adding locks to a building, but some of the emergency exits become slower to use.
HOST BAnd Hopper confidential computing had unencrypted NVLink, which compounds the security concern. So Nvidia is trying to solve both performance and trust at the same time.
HOST AThis is where the company’s ambition gets almost philosophical. It’s not enough to make the fastest chip. The market now wants the fastest chip that is secure, scalable, energy-aware, and easy to deploy.
HOST BAnd possibly portable across form factors. Which brings us back to the Windows SoC announcement. Nvidia clearly wants a piece of the AI PC story.
HOST AThat story is subtle. It’s not just “run a chatbot on your laptop.” It’s local inference, personal workflows, maybe lower latency, maybe privacy, maybe less cloud dependence.
HOST BAnd it also gives Nvidia a path into a market that is smaller per unit than datacenter but much broader in distribution. That can be strategically useful even if the margins are different.
HOST ANow, competitor moves. Qualcomm and Intel are the obvious reference points on AI PCs, because if Nvidia is shipping a Windows SoC, it is directly stepping into their neighborhood.
HOST BBut the more interesting competitive angle is that Nvidia is not just competing on the chip. It’s competing on the full developer experience, which is a very Nvidia thing.
HOST AAnd in models, Nemotron 3 Ultra puts it in a weird, interesting place relative to open-weight model makers. It’s not trying to be OpenAI, exactly, but it is trying to matter in the model layer.
HOST BMaybe the best way to say it is: Nvidia wants the model to be a proof point for the hardware, and the hardware to be a deployment path for the model.
HOST AThat’s a good formulation. And the humanoid robot reference design is another competitive signal, because it says Nvidia thinks physical AI will need standardized stacks too.
HOST BWhich means other robotics companies may not be competitors in the classic sense. Some may be customers, some may be partners, some may be building on Nvidia’s base layer.
HOST AAnd then there’s the optics ecosystem with Ayar Labs. If that category matures, it could quietly reshape who wins at the cluster level, because interconnect becomes a design advantage.
HOST BThat’s the thing: the competition around Nvidia often doesn’t look like a direct clone. It looks like a thousand little attempts to remove one layer of its advantage.
HOST ASo what should we watch over the next 30 to 90 days? First, whether the Windows SoC turns into a real product with shipping credibility, or stays mostly a strategic announcement.
HOST BSecond, whether Nemotron 3 Ultra gets actual benchmark data and real adoption, because right now the claim is interesting but incomplete.
HOST AThird, whether Rubin timing stays intact. If the Vera Rubin story starts slipping, that changes how people interpret the whole 2026 roadmap.
HOST BFourth, whether the confidential computing performance hit gets resolved in a way that doesn’t undercut the broader Blackwell narrative.
HOST AFifth, whether the NVLink Fusion and co-packaged optics ecosystem becomes real infrastructure or just a nice slide-deck story.
HOST BAnd sixth, whether the robot reference design becomes the start of a wider physical AI platform, or just a demo that made for good Computex photos.
HOST AI think the bigger falsifiable question is whether Nvidia can keep expanding without diluting itself. Because right now it’s trying to be datacenter, networking, models, PCs, and robotics all at once.
HOST B...and that could either be the greatest strategic moat in AI, or a sign that the company sees the core datacenter story maturing and is searching for the next layer of growth.
HOST AThat’s the thing I didn’t quite expect when we started: Nvidia doesn’t feel like a company only defending a lead. It feels like a company trying to redefine what the lead even means.
HOST BYeah. I came in thinking this would be about another giant chip cycle. But the more we walk through it, the more it looks like Nvidia is becoming infrastructure for multiple kinds of intelligence at once.
HOST AWhich is a very different kind of company than the one that started with gaming graphics in 1993.
HOST BAnd maybe that’s the real story. Nvidia didn’t just ride the AI wave. It turned itself into the place where the wave has to break.