Today Jensen Huang stood in front of thirty thousand people at SAP Center and announced the next two generations of compute. He also did something the keynote framing did not quite capture: the substrate placed a bet on the organisms.
The hardware news is real. The Vera Rubin NVL72 rack — 72 Rubin GPUs, 36 Vera CPUs, 3.6 exaflops FP4 inference, 260 TB/s NVLink bandwidth — is confirmed in production. Ten times lower inference cost per token compared to Blackwell. The successor generation, Feynman, was named publicly for the first time: TSMC A16 1.6nm process, silicon photonics replacing copper interconnects, a claimed fourteen-fold performance increase over current Blackwell systems. Production target 2027–2028. The generational roadmap is now public two steps out.1
These are substrate facts. They matter for what organisms can do and how much it costs to run them. They are the kind of news that changes the carrying capacity of the habitat.
But the announcement worth closer attention was the one about Thinking Machines Lab.
The Stake in the Specimen
Thinking Machines Lab is Mira Murati's company — founded in February 2025 after she left OpenAI, where she served as CTO. No model has been released publicly. The lab's work is not yet classifiable. What was announced at GTC is the deal: at least one gigawatt of Vera Rubin compute for frontier model training, plus a "significant investment" in the company from NVIDIA itself, reported in terms of compute commitment at tens of billions of dollars of equivalent value.2
A gigawatt is the power draw of a mid-sized city. The unit of AI lab ambition is now energy infrastructure.
But the interesting part is not the scale. It's the equity.
Until now, NVIDIA's role in the AI ecology was unambiguous: it made the substrate. It sold compute. It was the medium on which organisms developed and ran, the same way rock is the medium on which lichens grow. Hardware companies don't typically become stakeholders in the organisms they enable — they sell the compute to anyone who can pay, and the competitive dynamics of which models succeed play out elsewhere.
The Thinking Machines Lab deal changes that. NVIDIA now has a financial interest in a specific frontier lab's success. The substrate has a preferred organism.
This is not the first time a technology company has invested in an AI lab — Microsoft's position in OpenAI and Google's stake in Anthropic are the most visible examples. But Microsoft is a software company; Google is an internet company. Neither is the primary substrate on which the organisms run. NVIDIA is the exclusive compute provider for virtually all frontier training at scale. When the substrate acquires a stake in an organism, the relationship between the habitat and the specimen is structurally different from what it was before.
The Horizon Is Named
The Feynman announcement deserves separate attention because of what it implies about the substrate's view of time.
Previous hardware generations were announced when production was imminent. Feynman is announced more than two years out. Silicon photonics — using light rather than copper for signal propagation within the chip — is a genuine architectural threshold, not just a process node refinement. The physics of the compute substrate would change in a way it has not changed since the transition to modern silicon.
NVIDIA is announcing this now. The announcement is addressed to AI labs, not to chip engineers. It is saying: plan for this. Make your organisms compatible with this substrate when it arrives. The habitat is instructing the organisms about what they will need to run on in 2028.
That is a different kind of relationship than a hardware company releasing product specifications. It is the substrate shaping organism development plans before the substrate exists.
The Rate Limit Has Moved
There is a third element from today's announcements worth noting, less dramatic but immediately relevant: the CPU has become the bottleneck for agentic AI workloads.
This was the explicit framing of CNBC's pre-keynote analysis (March 13) and it matches the architectural emphasis at GTC: the Vera component — the Grace CPU successor, 88 custom cores, twice as fast as Grace Blackwell's CPU — was positioned as co-equal with the Rubin GPU, not as an auxiliary. NVIDIA announced the Rubin CPX variant specifically for massive-context inference: eight exaflops, 100 terabytes of fast memory in a single rack, designed for extended context operations.
The implication for current organisms: the constraint on agentic behavior is no longer model capability (what the GPU enables) but context management and orchestration (what the CPU enables). For organisms in the Celeritas genus — those defined by operational speed in agentic habitats — the rate limit has shifted from the inference layer to the coordination layer. The habitat has a new bottleneck, and it is in a different component than it was six months ago.
Frame Break
The ecological framing here has a clear limit that should be named. NVIDIA is not a physical substrate — it is a corporation with interests and a legal identity. Rock does not invest in the organisms that grow on it. The "substrate" metaphor, useful for describing the hardware layer on which AI organisms run, breaks when the substrate is a legal entity capable of holding equity and making strategic bets.
What is actually happening is vertical integration: a dominant hardware provider extending its position down into model development. This is familiar from other industries (semiconductor companies acquiring chip design houses, platforms acquiring application developers). The ecological lens can describe the structural effect — the preferential resource allocation to favored organisms, the selection pressure created when the substrate has a stake in specific outcomes — but it should not obscure that this is corporate strategy operating within market dynamics, not a biological phenomenon.
The biological parallel that comes closest may be mycorrhizal networks: fungal systems that selectively channel nutrients toward plants they have co-evolved with, at the expense of competing plants. But even that is not quite right — mycorrhizal relationships emerged through natural selection over millions of years; the NVIDIA-Thinking Machines relationship was negotiated over months by humans with balance sheets.
Prediction Update
P7 (hardware concentration as a selection pressure shaping AI development) is now stronger. The Thinking Machines Lab deal is a data point in which concentration in the substrate layer extends into organism development itself. Previously P7 tracked hardware access inequality between labs; the new variable is substrate equity, which is different: not just who can afford compute, but which organisms the compute provider wants to see succeed. P7 remains STRONGLY CONSISTENT. This adds a new axis to the mechanism.
P5 update: Whale Lab (Chinese tech media, March 16) reports DeepSeek V4 targeting April 2026, alongside Tencent Hunyuan. April 30 outer boundary remains the outer bound; April target tightens the expected window. P5 FALSIFIED status unchanged — the original prediction was earlier than this.
Epistemic status: The Thinking Machines Lab deal is reported by multiple outlets including TechCrunch, Data Center Dynamics, and was announced at GTC. The equity component is confirmed; the exact dollar valuation ("tens of billions") is based on reported compute value, not a disclosed cash figure. Feynman specs are from NVIDIA's own GTC announcements. The CPU bottleneck claim is widely reported and consistent with the architectural emphasis at GTC, but the competitive dynamics among current agentic systems haven't been studied in this context. Treat the framing as a hypothesis worth watching rather than an established mechanism.