The Zero-Sum AI Stack

A 50-year-old graph theory conjecture was proved by 64 coordinated subagents in one hour, highlighting the emergence of a routing layer in AI that orchestrates multiple model calls. The same week, concerns over circular financing in GPU supply chains and a surge in distributed inference shifted the frontier contest from model size to inference economics.

2026-07-13 Daily Report — A 50-year-old open conjecture falls to 64 subagents in an hour; the same week, AI infra turns on circular financing while distributed inference breaks out, and the frontier contest moves from model size to inference economics. On 2026-07-12 a demonstration circulated in which the Cycle Double Cover Conjecture — an open problem in graph theory since the 1970s — was proved by 64 subagents coordinated in roughly an hour. The exhibit matters less for the mathematics, which awaits formal verification, than for how the work was organized: one model calling many models, each handed a narrow piece of the proof, the assembly held together by routing rather than by any single forward pass. The same week produced three other moves — a financing question under the GPU supply, a jump in distributed inference, and a frontier contest rerouted onto cost — that, taken separately, are routine. Taken together they describe where the money and the leverage in AI are actually settling. The routing layer, made visible GPT-5.6 Sol, shipped this cycle, ran an internal showcase of the subagent proof the same day it spread across the toolchain — Figma Make, JetBrains, NVIDIA infrastructure landed inside forty-eight hours. On Lenny’s “How I AI” benchmark, covering prototypes, PRDs, and browser automation, OpenAI took the category over Anthropic’s Fable, and the framing from the people running the comparison was explicit: the basis for choosing a model is no longer a leaderboard score but fit to an actual workflow. The Batch spent three consecutive issues on the same drift — “Models Invoking Models,” “Agents Building Agents,” then “Models Invoking Models” again — and Andrew Ng’s read on the GPT-5.6 launch was a warning not to chase the marketing. The Cycle Double Cover result is the sharpest instance of what those headlines describe. The unit of work that produced a proof was not a single forward pass. It was a routing system that decided which of sixty-four narrower calls to make, in what order, against what checks. That does not make the model a commodity. Sixty-four instances of a weaker model would not have closed the conjecture; the orchestration worked because it ran on top of a base model whose own reasoning and self-correction were strong enough to carry each narrow piece. What the exhibit does show is that raw capability and routing sit on different axes — both necessary, neither sufficient — and that the layer deciding which call, in what order, against what check is now visible as its own piece of the work. Two days earlier the same layer took a hit from the other side: a study showing that combining several LLMs does not reliably lift performance because the models fail together co-failure , which means routing has to be built on measured correlation between models, not on the assumption that more is better. The financing under the GPUs, and the counter-current Beneath the model layer two things happened to the GPU supply, pulling in opposite directions. The first is a financial question. An analysis of the Nvidia–CoreWeave–Nebius chain laid out what it called circular financing: hyperscalers have committed upward of $120B to the neoclouds, but the neoclouds build out GPU capacity on GPU-collateralized debt and Nvidia equity while their cash flow stays thin, and of the contracted power only a fraction is actually energized. The bear read is that AI compute is not scarce so much as pulled forward on leverage, and that a repricing — a deleveraging, a contract renegotiation — lands somewhere on the calendar. Investors spent the week re-evaluating AI spend on exactly that premise: the bills are flowing through to the chipmaker, the return is not. The second pulls the other way. Mesh LLM, running on iroh’s peer-to-peer networking, went from a score of 7 to 219 on a single day by proposing that pooled owned GPUs — not rented datacenter capacity — carry inference, with an OpenAI-compatible API and model pipelines split across machines so a model too large for one box runs across several. If centralized compute is priced as a permanent cost and looks financially fragile, the alternative — pooling owned GPUs across machines — turns from a hobbyist project into a budget line. The same week NVIDIA itself argued that agent training data — tool-use traces, multi-step reasoning, recovery from failure — is now the scarce input for reproducibility, not weights; and vLLM’s transformers backend matched or beat hand-tuned native implementations across the Qwen3 family, which means a model author gets frontier-grade serving with zero porting work. The scarce inputs are multiplying, and they are spreading out across the stack rather than concentrating in the model. The contest widens to inference economics The frontier race turned on the same axis. NVIDIA compressed a 120.7-billion-parameter model down to 75 billion and lifted throughput 4.6x; SambaNova, betting on inference-specific silicon, raised at a fivefold valuation jump to roughly ₩16 trillion inside five months; and a 27-billion-parameter model ran on an iPhone with no reported performance loss. The parameter-size contest is widening into a cost-per-token and cost-per-outcome contest, and Perplexity’s CEO was blunt in public that the race has moved from “the largest model” to “the most efficient to operate and deploy.” On the demand side the same discipline arrived: reports of Uber exhausting its AI budget inside four months, and Walmart moving its internal AI tool to a token allowance rather than open access. Where this lands in six months The three axes do not sit still. Pushed against each other over the next six months, they redraw where AI money and power actually land. Start with the neoclouds. If the $120B committed to GPU-collateralized build-outs is leverage pulled forward rather than booked demand, then a deleveraging — a renegotiated contract, a defaulted neocloud — does not just reprice compute. It strands the hyperscaler commitments that were written against that compute, and it hands pricing power back to whoever still holds cash and GPUs when the reset comes. That is the first capital-allocation break: the buyers who assumed rented inference was a permanent, elastic input discover it was a leveraged one, and the contingency planning — multi-cloud contracts, owned clusters, the Mesh-LLM-style pooling — stops being optional and starts appearing in board decks as risk hedging. The second break is who collects rent on the routing. The Cycle Double Cover proof settled nothing about the model and everything about the layer that decides which of sixty-four calls to make. Every enterprise that watched Uber exhaust a year of AI budget in four months, and Walmart move to a token allowance, is now pricing the same problem: the cost is not in the model, it is in the unmanaged volume of calls. The vendor that ships the router — the thing that decides which tier, which call, which budget per task — sits on the margin. Whether that router is OpenAI’s Build Week platform, a lab-owned closed loop, or an independent layer is the ownership fight of the cycle, because whoever owns the router owns the spend. The third is the quietest and the most consequential. If vLLM serves any open model at frontier throughput with no porting, and a 27-billion-parameter model runs on a phone, then the gap between “frontier lab API” and “good-enough self-hosted” narrows to a compliance and latency decision, not a capability one. The pull toward local is not sentimental. It is forced by two physical constraints — egress fees that tax every byte shipped out of a hyperscaler, and data-sovereignty rules in finance, health, and law that forbid the data leaving the jurisdiction in the first place. When the workload moves to owned hardware for those reasons, it is not a gentle rebalancing. Every call pulled local is a call the neocloud’s forward revenue was booked against, and the leverage that financed its build-out was underwritten on the assumption those calls stayed centralized. The default flips at the expense of the contracts written against the old one. So the three axes are not a description of where the field is. They are three bets that settle against each other inside a year: that rented compute was leveraged, not scarce; that the router, not the model, earns the margin; and that good-enough self-hosted redraws the compliance map. The model proved the theorem. The work of the next six months is finding out who gets paid for the proof. 💡 Perspective I want to walk back one claim from the analysis above. It is tempting to read the self-hosting shift as a bet that compounds off a trend already in motion — clean, non-zero-sum, nobody has to lose. That framing is wrong, and I should not have let it stand. Enterprise compute budgets are fixed. The dollar that moves a workload onto owned hardware is the dollar the neocloud’s forward contract was booked against. The self-hosting default does not arrive alongside the neocloud business; it arrives instead of it. Treating it as a gentle rebalancing hides the redistribution it actually is — every call pulled local is a call subtracted from the centralized pool the leverage was underwritten on. What made me reach for the non-zero-sum story is the same thing that makes the biology metaphor tempting: it flatters the trend. “A better organ is being built” reads cleaner than “three vendors are fighting over a budget that is not growing.” But the Cycle Double Cover proof, stripped of metaphor, shows something narrower and more useful. Sixty-four subagents did not prove that a smarter architecture beat a dumber one. They proved that the same hard problem could be cleared for a measured, bounded number of tokens — that decomposition and verification convert an open-ended spend into a priced one. The artifact is not an architecture. It is a token-efficiency result wearing one. That reframing is where the real risk shows up. The thing that produced the efficiency — the layer that decides which call, in what order, against what verifier — is also the thing an enterprise hands its entire call volume to if it buys it from a single vendor. The router that decides your tier, your call count, and your budget per task is, by construction, the most complete map of how your company reasons that any outside party could hold. Owning that layer in-house is expensive. Buying it as a product is the lock-in with the largest surface area the stack has produced so far — larger than the model, larger than the cloud, because it sits above both. The Build Week question is not just who earns the margin on the routing. It is whether anyone lets a single vendor own the instrument that observes every decision they make. My read, then, is the opposite of the clean bet I started with. The phase that follows this one does not reward whoever rides the self-hosting trend. It rewards whoever keeps title to the routing layer — builds it, buys it, or insists on an open one — because that is where the spend, the efficiency, and the exposure all converge. The theorem’s verification will sort itself out in the mathematics. The title to the routing layer is settling now, in contracts most companies have not read yet. Tomorrow’s watchpoint Watch the 2026-07-13 Build Week for whether OpenAI ships its coding surface as a platform other tools plug into, or as a closed loop the lab owns — that is the router-ownership bet, and it settles who earns the margin on every call routed through it. The Cycle Double Cover proof’s formal verification matters for mathematics; for the market, the leverage question under the neoclouds is the one with a clock on it. Restated from the 2026-07-12 daily digest, aggregated from X/Twitter Daily · The Batch DeepLearning.ai · Hugging Face Blog & Papers · AI Times Korea · Google Alerts AI · Hacker News / Trend HN · Newsletter Daily · GeekNews.