Eric Anderson ∞/21M

$IREN $NBIS $NVDA $MRVL I have been in the markets long enough to know that the next few sessions will be all about FUD and the market trying to find its footing regarding AI. Let me start by saying that the Amodei is not this era's industrialist. If at all, he is more like the Orren Boyle, fails to deliver due to lack of production capacity and retorts to washington to get his way through politics. Nice, but the Hank Rearden of this age is Jensen. On the same token, I do not think Washington will give him what he is asking for. And I do not think Jensen has spent years building the neocloud ecosystem just to hand the market back to four hyperscalers. The equilibrium is more likely to be a regulated race. Faster AI development behind more checkpoints, with a certified layer of independent infrastructure carrying much of the buildout. Start with Washington. The US currently leads the technology. China remains close enough that compliance with any global slowdown would be difficult to verify. AI investment is also doing real macro work at home. The Federal Reserve estimates that AI-related investment contributed about 1.36 percentage points to US GDP growth in Q1 2026. That investment is not yet the deflationary dividend. In the short run it raises demand for capital, power, equipment and labor, potentially lifting the neutral rate. The dividend comes later, if cheaper intelligence raises productivity and reduces unit costs across the economy. Slowing now risks removing the investment impulse before the productivity payoff arrives. That is a difficult proposition for any administration facing China, a heavy fiscal burden and an economy leaning on AI capital formation for growth. The White House’s revealed policy already points elsewhere. It wants classified capability testing, government access before sensitive releases and stronger security. But it has explicitly rejected mandatory licensing or preclearance of model development, while directing the national security system to obtain the most advanced models without delay. That is acceleration with guardrails. Now look at NVIDIA. If hyperscalers become the only entities allowed to own and operate frontier compute, NVIDIA loses strategically even if it keeps selling chips. Microsoft, Amazon, Google and Meta would control access to the customer, set the rental price, bargain against NVIDIA at enormous scale and keep funding their own accelerators. The long-term result would be fewer buyers, greater pricing pressure and a faster migration toward custom silicon. Neoclouds are Jensen’s answer to that concentration. NVIDIA invested $2 billion in CoreWeave. It is using its financial strength to help CoreWeave secure land, power and shell. It is testing CoreWeave software for inclusion in NVIDIA reference architectures. It has brought IREN and other independent operators into Australia’s 2 GW sovereign-compute buildout. And it is working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than $500 billion of third-party capital for AI factories. That is not a temporary distribution channel waiting to be absorbed by hyperscalers. Far from it. It is an attempt to turn NVIDIA compute into an independently financed asset class, widen the buyer base beyond four companies and make CUDA the architecture through which labs, enterprises and governments access intelligence. Jensen can accept Amodei’s safety diagnosis without accepting his industrial solution. His preferred version will be a distributed but certifiable ecosystem. The labs build models. NVIDIA defines the common architecture. Neoclouds and sovereign operators provide capacity. Governments evaluate capabilities, control sensitive deployment and audit the infrastructure. Safety becomes the standard that lets the ecosystem remain distributed, not the reason to collapse it back into hyperscalers. That produces a very different second-order effect for neoclouds. The sector does not disappear. It bifurcates. Baremetal GPU renters with no security, orchestration or customer layer become weaker. Certified operators able to protect model weights, isolate workloads, document access, satisfy sovereign requirements and redeploy capacity between customers become more valuable. That is why the software layer matters. $NBIS owning inference optimization and orchestration becomes more important than simply owning accelerators. $IREN’s power portfolio gets it into the game, but Mirantis, customer access and workload portability determine whether it becomes a trusted platform or remains a landlord. CoreWeave’s relationship with NVIDIA shows what the integrated version can look like. The most important regulatory detail is where Washington attaches liability. If the model developer remains responsible for model behavior, frontier labs can outsource physical capacity to certified neoclouds. That is bullish for the strongest independent operators. If the compute provider becomes jointly responsible for what every model does, the compliance burden may become so large that only hyperscalers can absorb it. The physical assets survive, but independent equity economics get squeezed and more of the value migrates upward. There is also a third-order geopolitical effect. Export controls plus sovereign AI requirements fragment the world into trusted compute blocs. Governments duplicate capacity because they do not want critical intelligence running entirely inside another jurisdiction. The frontier can become more regulated while the physical buildout becomes larger and less globally efficient. So I would not read Amodei’s essay as the beginning of the end of AI infrastructure. I know many will see it that way, but I thin it is the opening argument in a fight over who gets licensed to build, finance and operate it. Amodei wants more time. Washington wants technological overmatch and the eventual productivity dividend. Jensen wants enough independent operators to prevent the hyperscalers from owning the entire market. The compromise is likely to be be certified acceleration.