The Market Is Underestimating This Massive AI Demand Shift

The fear that dominated the market for weeks was that AI was slowing. Jose Najarro answers with the opposite picture: in just the last week, Grok 4.1, Claude Opus 5.5 and OpenAI's GPT-6 Luna arrived back-to-back. His read is that when model teams talked about 'speed,' they meant adding safety layers, not braking innovation; if an extra step makes a smarter model, demand keeps compounding.

The hard signal that backs him is the reversal in rental pricing. Citing Small Cap Snipers, Najarro notes Nvidia's B200 on-demand rental hit $7.88 per GPU-hour, a fresh high. With Blackwell's 180 GB HBM3e cards trading at a $6.2-$6.5 median for months, that print suggests compute supply is tight and pricing power has shifted to sellers.

Behind the price move is the agentic loop. Viral uses of Muse — saving money, doing a specific job end-to-end — lift not only GPU demand but CPU-heavy orchestration: code execution, tool calls, sandboxing, analytics and data pipelines outside the model. The agent must be smart, so the model must be smart; one lift does not crowd out the other, they climb together.

That is why Nvidia's Vera bet matters. The chip is shipping this quarter to AWS, Oracle, OpenAI and Anthropic, among others. On the last earnings call, Jensen Huang guided to more than $20 billion in standalone Vera sales this year — not counting Rubin combos. For a product entering the CPU market in the second half, that scale is a category created inside a single half-year.

Nvidia trades around $228 with a $5.5 trillion market cap, and Najarro flips the 'how much higher?' question. He remains overweight Nvidia for the long run and sees Vera as a second-half lever that can push results above expectations. AMD, Intel and Qualcomm can also benefit, but the best position belongs to whoever builds the full stack.

The full-stack point leads straight to his neo-cloud thesis. In its latest release, Nvidia named Cloudflare, CoreWeave, Nebius and Oracle Cloud Infrastructure as Vera system builders, with AWS planning a pure-CPU build as well. What the market may not have priced is the multiplier on near-term rental revenue for those providers.

The most idiosyncratic near-term trace is Akamai. The stock is up about 39% year-to-date yet still about 32% below its 52-week high. Najarro does not pitch it as an investment, only as an idea to watch: if Akamai can pair its edge distribution network with CPU offers for agentic demand, it could find pricing flexibility — but he frames it as a watchlist notion, not a call.

The more structural story is in large deployments. Meta is building AI infrastructure with Nvidia around wide Vera rollouts. SpaceX disclosed on August 24 that it adopted Vera CPUs to accelerate its H100K-scale work. As H100K tooling and workloads expand, anyone with sizable CPU capacity rides the wave, which is why CoreWeave's recent partnerships stand out.

At CoreWeave, price and financing are strengthening together. On September 17 the company said it is contracting new compute at higher prices: short-duration deals of 3 to 6 months are printing around $40 million per megawatt. In parallel it upsized a $3.7 billion convertible senior notes offering and added roughly $3 billion via an at-the-market raise, taking total proceeds above $5 billion to fund more data centers while GPU prices keep climbing.

His second-largest neo-cloud holding is IREN, with Nebius as the closing idea. IREN sits at a $14-20 billion market cap, stands out for locked-in power to hedge the 2026-2028 energy crunch, and leans on its Nvidia tie — management pay criticism does not change his market read. At Nebius, the CEO frames spot and dynamic pricing as a programmatic auction, arguing that even inside infrastructure — which he calls the least glamorous part of AI's most glamorous sector — there is pricing innovation. Nebius is up 162% year-to-date, IREN added 14% in the last five days, and a fresh Palantir partnership alongside the Palantir-Nvidia link underpins the momentum.