AlphaZero mastered chess in four hours by playing itself. No human games, no grandmaster data, just the rules and compute. It surpassed all human knowledge of chess in an afternoon. Silver is betting that approach generalises into science, medicine, and engineering. If it does, the LLMs generating $30 billion a year in revenue are the equivalent of training a chess engine on grandmaster games when you could let it play itself. LLMs will still exist, but the question is whether they hit a ceiling that reinforcement learning does not.

The AI industry's three scaling laws all describe ways to make LLMs better: pretrain on more human data, refine with post-training techniques like distillation and RLHF, or apply more compute at inference for chain-of-thought reasoning. Every GPU cluster in this issue is optimised for one of those three relationships. Silver's work implies a fourth: experience scaling, where performance improves as the agent interacts with a simulated environment more, generating its own training signal through trial and error. The compute requirements are real (AlphaZero used thousands of TPUs for self-play), but the workload is fundamentally different. LLM training is one massive parallel job across the cluster. RL training is millions of small sequential loops, each one latency-sensitive, and bottlenecked by how fast you can simulate the environment rather than how many GPUs you can rack. If experience scaling works for real-world problems, the winning infrastructure shifts from "biggest cluster" to "fastest simulation," and the hundreds of billions committed to LLM-optimised clusters may be aimed at the wrong scaling relationship.

Sequoia leading this round is the signal that matters. The most successful venture firm in AI history is hedging against the very buildout it helped fund. That is not a research grant or a government initiative. That is $1.1 billion of private capital placed behind the one researcher most qualified to test whether LLMs are the destination or just the first stop.

This is the first time Beijing has ordered a completed US tech acquisition unwound. Last week the White House accused China of industrial-scale distillation of US frontier AI (Issue #103). This week Beijing ordered Meta to return a Chinese AI startup. Both governments are now blocking the other's access to AI intellectual property simultaneously, and the companies caught between them are the ones getting crushed.

"Singapore washing" just became a regulatory risk factor for every Chinese AI startup with Western investors. Manus moved to Singapore specifically to access foreign capital and sidestep both countries' restrictions. Beijing reached through the Singapore incorporation and ordered the deal unwound anyway, asserting jurisdiction based on the founders' nationality, the technology's origin, and the original R&D location rather than the place of incorporation.

Losing Manus does not kill Meta's agent strategy. Muse Spark, the Broadcom MTIA partnership, the CoreWeave $21B agreement, and data centre #32 in Tulsa are all intact. But it removes the team behind Manus's agent tooling, and it sends a warning to every hyperscaler considering acquisitions of Chinese-origin AI talent, regardless of where that talent has incorporated.

We have tracked Sharon AI across the catalogue. In Issue #100, they had 432 GPUs online and a Canva contract. In Issue #102, we contrasted them with NewBird AI as proof that small starts are not disqualifying if the infrastructure DNA is right. Six months later, Oaktree is leading $350 million in convertible notes. The capital structure has outpaced what most observers thought possible for a company at this scale.

Oaktree manages $190 billion and is one of the largest credit investors in the world. Its involvement means the credit market is now underwriting GPU infrastructure at every scale tier, not just at CoreWeave ($8.5B A3-rated debt) or Firmus ($10B Blackstone facility) level. The asset class is institutionalising from the top down and the bottom up simultaneously.

Compare this to NewBird AI: $50 million, no named hardware, no named counsel. Both are NASDAQ-listed GPU plays. The substance between them is separated by two orders of magnitude, and the market is starting to tell the difference. Oaktree is the kind of investor that does not write cheques based on narratives. It writes them based on contracted revenue and asset value.

The strategy is straightforward: NVIDIA does the expensive work of training frontier-class models on massive GPU clusters, then gives them away for free. Enterprises download the model, fine-tune it on their own proprietary data using NVIDIA's NeMo tools running on NVIDIA GPUs, and deploy on NVIDIA inference hardware. No API fees, no per-token charges, and the enterprise's data never leaves their own infrastructure. Every download is a customer who needs NVIDIA hardware to fine-tune and serve it.

This is the data sovereignty problem solved from the hardware side. Anthropic solves it from the cloud with Cowork on 3P (Issue #103). Google solves it with confidential Blackwell VMs. NVIDIA solves it by putting the model in your hands and letting you own everything. For defence, healthcare, and finance, where sending prompts to a third-party API is a compliance problem, owned hardware running open weights is the path of least regulatory resistance.

Nano Omni is not a chat model. It is a perception sub-agent designed for swarms, functioning as the "eyes and ears" alongside larger models that handle reasoning and execution. That means more data in, better decisions out. In theory, at least.

The AI data centre silicon market just added a fifth vendor. NVIDIA dominates. AMD has Meta's 6GW deal (Issue #93). Broadcom builds custom silicon for Meta (MTIA) and Google (TPUs). Amazon designs Trainium and Graviton in-house. Qualcomm entering with a hyperscaler win means there is now a mobile-first chip designer competing for inference workloads alongside the incumbents, and the unnamed customer is large enough to move Qualcomm's stock 15% on the announcement alone.

Qualcomm's Arm-based architecture is optimised for power efficiency, which is the binding constraint in every data centre story in this issue. The floating data centre in Singapore, the rejected facility in Minnesota, the $4 billion project that died over zoning are all fundamentally about power. A chip designed for mobile power envelopes entering data centres at hyperscaler scale suggests inference workloads are fragmenting across silicon architectures faster than the market expected.

"The first of many" from the CFO is the signal to watch. If Qualcomm lands a second hyperscaler before analyst day, the competitive dynamics in AI inference silicon shift significantly. NVIDIA's inference pricing power depends on being the only option at scale. Every new silicon vendor that wins a hyperscaler contract compresses that margin.

The hyperscalers are absorbing cost increases that would break smaller operators. DRAM up 158%. H200 spot up 68%. Oil prices spiking from the Iran conflict. Helium shortages hitting semiconductor manufacturing. And yet AWS is not raising prices, Microsoft is not raising cloud prices, and none of the three have revised capex guidance downward. The message: demand for AI compute is strong enough that the hyperscalers will eat hundreds of billions in cost increases rather than risk slowing adoption by passing them through.

Ornn's GPU pricing data is now being cited by CNBC as the reference for H200 spot markets. That price signal matters because it is the closest thing the market has to a real-time clearing price for GPU compute. The 68% increase from $2.27 to $3.82 per hour since January tells you that GPU scarcity is intensifying even as new capacity comes online. Supply is growing. Demand is growing faster.

$500 billion in collective hyperscaler AI capex for 2026 is the number that makes the Ineffable thesis so pointed. That is half a trillion dollars committed to LLM-era infrastructure in a single year. If experience scaling works for real-world problems, some portion of that spend is optimised for the wrong workload. If it does not, the hyperscalers have correctly priced the largest infrastructure buildout in history. Both outcomes are possible. Neither can be unwound.

Google's India groundbreaking shows that capital is flowing to jurisdictions that say yes. India wants the investment, has the land, and has two construction partners ready to build. Meanwhile Virginia, the largest data centre market in the world, just lost another project to regulatory barriers. Utah residents found out about a facility that would use more power than their entire state through a Facebook post the morning of the vote. The industry's social licence is eroding in the jurisdictions where it needs it most.

The Utah story is the most extreme version of the permission problem we have covered. MIDA, a military infrastructure authority, used its special powers to bypass normal planning processes for a 40,000-acre project backed by a celebrity investor. Residents had no notice. The county commission was prepared to vote the same day. Community opposition stopped it, but the fact that a project of this scale nearly reached final approval without public awareness shows how badly the siting process has broken down on both sides.

Compass in Virginia and Google in India are the same story told in opposite directions. One jurisdiction's regulatory friction is another's competitive advantage. As US permitting timelines stretch and community resistance intensifies, the gigawatt-scale projects that define the next phase of AI infrastructure are increasingly being sited in countries where approval is faster and opposition is weaker. The capex follows the permit.