The Read

The constraint on the AI buildout acquired a share price, and the market priced it unevenly. Neocloud equities took the damage — IREN down 5.93%, CoreWeave down 5.41% and Galaxy Digital down 4.25% in Wednesday's early session, capping a month in which CoreWeave, IREN and Nebius shed 38% to 45% — while hyperscalers moved between 0.4% and 1%. The mechanism is permits rather than policy: the governors who won on Tuesday hold the approvals, Texas alone is 27% of planned US data-centre power, and Ben Thompson wrote the ladder out plainly — compute, then memory, then power, then politics, and politics is the rung you cannot buy. At the other end of the stack the compounding did not pause: Moonshot put 2.8 trillion parameters of open weights on Hugging Face twelve days after announcing the model, Broadcom's TPU compute output doubled in a single quarter to roughly two million H100-equivalents, and a physics benchmark built by 42 working physicists found the best of 26 agent configurations clearing 46 of its 168 runs. Concrete is getting harder to pour while intelligence gets cheaper, better and better measured — which is the day-zero trade stated as a spread, because the returns were never in the concrete.

🌊 Tide

No shift. Governance-as-market-structure took its second consecutive confirmation, and this one came off the tape rather than the ballot box: Tuesday's midterm results said what voters want, Wednesday's price action said what it costs. The new information is not that the buildout faces political friction — the datacenter-backlash wave has logged that for months — but that the friction now sorts the AI trade into two books, and the sorting key is who needs a new permit. Cost-collapse held, with its open-weight leg moving in the waves below. Distribution-rewrite and ai-as-worker were quiet.

The permitting constraint got a price, and it split the AI trade in two

Wednesday was the first session to price Tuesday's result, and it did not price it as a sector event. In the early session IREN fell 5.93% to $23.01, CoreWeave 5.41% to $50.83, Galaxy Digital 4.25% and Applied Digital 2.81%, against hyperscalers down 0.4% to 1% — and that is on top of a month in which CoreWeave lost 43.28%, IREN 44.9% and Nebius 38.67%. The dividing line is not AI exposure, it is permit exposure: Texas accounts for 27% of planned US data-centre power on Wells Fargo numbers reported by BigGo Finance via 24/7 Wall St., Georgia 7%, and Ohio and Pennsylvania 5% each, and the incoming governors hold the approvals that greenfield capacity needs. Ben Thompson put the sequence in order — the binding constraint moved from compute to memory to power and now to politics, which is the one you cannot solve with capital — and noted the corollary that hyperscalers gain share precisely because they hold more political sway and more existing sites they can repurpose or rebuild. Worth keeping the magnitude honest: the 27% is Texas's share of the planned pipeline, not capacity anybody has cancelled, and Eurasia Group's Heather West reads the new rules as adding "friction" while "it doesn't stop the buildout." Evercore ISI framed it the same way, as "incrementally more negative sentiment" rather than a thesis break.

So what: The cost of compute just split by provenance, and that is a procurement opening rather than a problem. Capacity already permitted and energised is now structurally scarcer than capacity on a drawing board, so anything you plan to run in 2027 is worth contracting against built footprint now, while the people who have to go get a permit are the ones repricing. The second-order read matters more for roadmaps: if political permission is the binding constraint, efficiency per watt stops being an engineering nicety and becomes the thing that decides who ships — which favours routing, distillation and open weights over brute capacity, every one of which you can buy today without asking a county for anything.

Sources: 24/7 Wall St. — Texas Blue Wave Could Threaten 27% Of Planned U.S. Data Center Power. Here's What It Means For CoreWeave, Iren, Nebius And Other Neocloud Stocks · Stratechery — AI on the Ballot · Yahoo Finance — AI data center stocks drop as Democrats sweep U.S. midterms · Euronews — US midterms: Data centre backlash helps Democrats win in Texas, Georgia and Ohio. What happens now?

🌊 Waves

Moonshot put 2.8 trillion parameters in the open, twelve days after announcing the model

Moonshot released the weights for Kimi K3.1 Pro on Hugging Face on Wednesday: 2.8 trillion total parameters, 104 billion activated, a 1,048,576-token context window, quantisation-aware trained in MXFP4, and 1.56 TB across 96 safetensors shards. Third-party trackers call it the largest open-weight release to date — Moonshot's own card does not claim it — and the interval that matters is twelve days — this brief covered the model's announcement on 23 October, when the weights had not yet shipped. Reports differ by a day on that announcement, so call the gap just under two weeks. On Moonshot's own card it posts Terminal-Bench 4.0 at 65.7 against Claude Opus 5.5's 66.4 and GPT-6 Astra's 58.2, DeepSWE v1.1 at 75.5 against 74.2 and 74.1 though below DeepSeek-V4.1-Flash's 77.0, BrowseComp at 92.1 against 90.8 and 91.5, and HLE-Full with tools at 70.0 against 67.7. Two honest qualifiers, each worth one clause: these are vendor numbers at maximum reasoning effort, and the card itself shows where the model is not close — HLE-Diamond 39.4 against Astra's 65.8, ExploitBench 74.7 against Astra's 100. And "open weights" is not open source here: the licence requires a separate Moonshot agreement for third-party serving above $20M of group revenue over any twelve months, and UI attribution above 100M MAU.

Roadmap implication: Plan on a weeks-long lag between the closed frontier and a downloadable equivalent, not a quarters-long one, and build the switch now rather than when you need it. The practical move is to make model identity a configuration value on every high-volume path this quarter, so the next release of this kind is a config change instead of a project — that is what turned Revolut's 90% and LinkedIn's 67% into realised savings rather than press releases. Read the licence before you plan a product on it: at this size the constraint on open weights is commercial terms and serving economics, not capability.

Sources: Hugging Face — moonshotai/Kimi-K3.1-Pro

Open-model inference finally got a denominator

Nathan Lambert published his attempt to size the open-model economy on Wednesday, and the useful part is that he treats it as a measurement problem rather than a talking point. The hard number: Fireworks announced at its first company conference that it is serving more than 40 trillion tokens a day on customers' customised models alone — one slice of its inference, and a vertical that has more than doubled since the summer. Lambert expects the firm's total to be well above 100 trillion tokens a day by his next update. His method uses Baseten, Fireworks and Together AI as bellwethers plus open-model share of OpenRouter volume, checked against the platform and API token volumes Google reports in earnings, and he states his own limits plainly: Bedrock and Azure do not break out usage by model category, so the startups are a proxy, and startup marketing metrics are unreliable. The closed-versus-open conclusion sits behind his paywall, so treat the shape of the curve as the finding and not the share.

Roadmap implication: This is the measurement layer arriving under a tide this brief has been calling for months, and measurement is what makes a bet contractable. If you are negotiating inference capacity or an open-model support contract, you now have public reference volumes to anchor against instead of a vendor's adjectives. The strategic read for anyone building a serving business is less comfortable and worth hearing: Lambert's own view is that the firms in the middle are squeezed between the labs they serve and the clouds they buy from. The margin is not in moving tokens — it is in the work the tokens complete, which is exactly where the gap between throughput and delivered value still sits wide open.

Sources: Interconnects — How big is the open model economy today?

Broadcom's TPU compute output doubled in a single quarter

Epoch AI published the numbers on Wednesday: Broadcom's TPU unit shipments rose about 65% quarter over quarter to roughly 800,000 chips in fiscal Q3 2026, while compute output nearly doubled to about 1.97 million H100-equivalents — against roughly 327,000 in fiscal Q1 2025. The gap between a 65% unit increase and a doubling of compute is the mix shift: TPU v6e supplied most TPU compute through fiscal Q4 2025, TPU v7 is now the vast majority of shipments at about 2.5x v6e, and v8i entered production in fiscal Q3 2026 at roughly 2.2x v7. The three preceding quarters grew between 2% and 22%, so this is a step rather than a trend line. Broadcom oversees nearly all TPU production today, with MediaTek expected to ramp by the end of 2026 and into 2027.

Roadmap implication: The silicon is accelerating on exactly the quarter the permits tightened, and that divergence is the arbitrage. Non-NVIDIA accelerator supply roughly doubling in capability terms is what makes the routing and efficiency work above economically real rather than aspirational — and it lands inside hyperscaler footprints that already have their power and their approvals. If you are planning 2027 capacity, the useful question is no longer how many chips exist but whose land they sit on. Price TPU-backed capacity against GPU-backed capacity this quarter; the two are diverging on political risk, not just on performance per dollar.

Sources: Epoch AI — Broadcom TPU compute shipments doubled from Q2 to Q3 2026, reaching 2 million H100-equivalents

Forty-two physicists built the test, and the best agent cleared 27% of its runs

FrontierPhysics went up on Wednesday: 56 tasks drawn by 42 physicists from their own research across nine domains of theoretical, computational and experimental physics, requiring an agent to run literature review, research plan, experiments, structured outputs and a manuscript, scored by deterministic verifiers plus expert rubrics for publishability rather than correctness alone. Across 22 models in 26 configurations, no configuration passed more than 46 of its 168 rollouts — a ceiling around 27%. The finding that matters more than the number came out of manual trajectory review: the binding weaknesses are insufficiently rigorous reasoning and limited expert intuition, not gaps in physics knowledge and not execution errors. Set it beside SemiAnalysis's GDPx from Monday, where the best model in the world scored 12% on eleven domains of real economic work, and the two independent reads agree on both the altitude and the failure mode.

Roadmap implication: A 27% ceiling on expert research work is not a warning, it is a measurement of how much room is left — and the failure mode tells you where to stand. If the gap is rigour and taste rather than knowledge or tool use, then the value accrues to whoever supplies the judgement layer: verification harnesses, expert review in the loop, and task decomposition that puts the human where intuition is scarce. That is a buildable product and a hiring plan, not a wait-for-the-next-model plan. It also sets the honest expectation for a board: agents are credible at the execution tier of expert work today and not at the judgement tier, so staff the judgement tier and let the execution tier compound.

Sources: Hugging Face Papers — FrontierPhysics: Evaluating Agents for End-to-End Physics Research

🌊 Ripples

Tencent shipped Hy4 with a weights link that goes nowhere

Tencent's Hy4 went live on Wednesday with a 1.0M-token context window and list pricing of $0.834 per million input tokens and $2.501 per million output, currently discounted 50% to $0.417 and $1.251, with cache reads at $0.042 list. It is the first Hunyuan model that can see, adding visual understanding alongside gains Tencent claims on agentic and long-horizon productivity work over Hy4 Preview; the scores carried on the model page include 81.0% on Vibe Code Bench v1.1. The notable part is what is missing: the model page points at Hugging Face for weights and that repository does not exist — huggingface.co/tencent/Hy4 returns a 404, and the 780-billion-parameter tencent/Hy4-preview that is on the Hub is a different model, last touched in August. An open-weight release is not open until the weights are there.

Do this now: Price it as a closed API today, because that is what it is — and that is still interesting at $0.834 in and $2.501 out for a million-token context with vision. Run your own agentic evaluation before you believe the page: these are vendor-published scores, and the gap between an announced weights link and a live one is exactly where procurement risk hides.

Sources: OpenRouter — Tencent: Hy4

A paper asked what multi-turn RL actually buys, and the answer is not what gets reported

A paper from Kejian Zhu and colleagues posted on Wednesday evaluates agentic reinforcement learning through Pass@(K,T), separating sampling breadth from interaction depth, and reaches two findings that cut against how these gains are usually presented. Trajectory-level RL advantages depend on a sufficient turn budget and shrink noticeably when the budget tightens; and agentic RL does not uniformly improve single-step decision quality — its gains concentrate in error recovery and long-horizon resilience. The authors propose BudgetMix RL, varying interaction budgets during training, and report it holds up across execution settings.

Do this now: When a vendor shows you an agentic RL improvement, ask what turn budget produced it and what the number looks like at the budget you can afford to run in production. If the gain is really error recovery rather than better decisions, that changes what you should measure — track recovery rate and completion under a capped budget, not per-step accuracy.

Sources: Hugging Face Papers — What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents?

Epoch rebuilt its capability index to say how sure it is

Epoch AI also moved the Epoch Capabilities Index onto a Bayesian model on Wednesday, replacing a frequentist best fit and 500 bootstrap resamples with posterior sampling by MCMC, a beta likelihood so predictions stay inside 0–100%, and a per-benchmark noise parameter instead of one shared value — which turns out to range from 0.038 on METR Time Horizons to 0.176 on VPCT, against the 0.076 previously applied to everything. For the 176 models at or above ECI 125 the estimates barely move, 0.9 points on average, while the 90% intervals narrow 13%; the 76 models below 125 all move down by 13.7 points on average and their intervals widen by 33%, which is the honest other half of the result. Out-of-sample 90% interval coverage improves from 84.2% to 89.6%. The output that is new in kind rather than degree: the index can now state that GPT-6 Astra is 72% likely to be more capable than Claude Opus 5.5, and over 99% against Claude Fable 5.1.

Do this now: A leaderboard that reports a probability instead of a rank is usable in a decision memo, and a 72% edge is a materially different fact from a one-place lead. Where a model choice is close, stop paying for the top of the table and start pricing the overlap — and note that benchmark noise differs by more than 4x across benchmarks, so a single-benchmark win is weaker evidence than it looks.

Sources: Epoch AI — Updating the Epoch Capabilities Index methodology to use a Bayesian model

The constraint moved to politics, so Musk started pricing orbit

Within hours of the result Elon Musk posted that "data centers in space looking pretty appealing right now," having given $474 million to Republicans in these midterms. It reads as a quip, and Ben Thompson's Wednesday list of implications arrives at orbital datacentres independently, by the same reasoning. The grounding numbers under the mood are worth keeping: an NBC exit poll had 56% of respondents calling AI and data centres a "crucial or a major factor" in their vote, 23% of them "crucial", and Data Center Watch counted at least 45 projects worth $68 billion blocked or delayed by local opposition in the second quarter of 2026 alone.

Do this now: Treat orbital datacentres as a signal about terrestrial scarcity rather than a 2027 procurement option. The actionable version of the same instinct is cheaper and available now: site new capacity where the politics are settled, and put the engineering budget into efficiency per watt, which pays off under every version of this constraint.

Sources: Euronews — US midterms: Data centre backlash helps Democrats win in Texas, Georgia and Ohio. What happens now? · Stratechery — AI on the Ballot · Insurance Journal — New Data Centers Worth $68 Billion Disrupted in US, Data Show

The full edition and the archive: https://excelsiorgroup.ai/insights/signal/