Predictions

Reported internal lean toward 2027, and $600M of unsold secondary inventory says the private market has not cleared at the current mark. You do not go public into that.

The reported October Nasdaq listing needs a public prospectus with road-show lead time, and the late-August timetable CNBC's sources gave has already slipped once, which is what our previous call caught. SB Energy's public filing this morning says the AI-adjacent IPO window is opening rather than closing, and a lab that wants to price before OpenAI has no reason to sit on the flip past this month. The call fails if the listing itself is drifting toward 2027, which would be the more interesting finding.

Every public signal into the summit points to a joint statement, not a treaty: Bessent's open-to-discussing framing, the UN's contact-not-treaty ask, and experts' hotline proposal all describe talks rather than instruments. A state dinner with eight tech CEOs reliably generates photographs and coverage; it does not reliably generate enforceable text, especially with chip export controls and the tariff truce competing for the same room.

August's information-sector loss of 23,000 jobs ran at nearly triple its own 12-month average, the fastest pace of cuts on record for the category, arriving the same month software, media, and AI labs kept shipping tools explicitly aimed at replacing that work. One month of a record is a data point. A second consecutive decline is the beginning of a trend a single print cannot be.

AIR disclosed privately in June and Microsoft shipped nothing in three months. Public disclosure changes the calculus: the exploit recipe is published, Anthropic and OpenAI have both patched, and an unpatched zero-click RCE in a flagship developer product costs more reputation every day it stays open. Vendors move when silence costs more than the fix.

DeepSeek built its identity on being the price floor of frontier inference, and it is now tripling V4-Pro peak pricing while telling developers demand exceeds capacity. If the cheapest capable lab holds a 3x increase, open-weight availability does not discipline API pricing the way the market assumes; if it cannot, the floor reasserts itself within a quarter. Either answer reprices every inference business model, and T3 runs the same experiment on Anthropic from the premium end.

CXMT crowned itself China's most valuable listed company seventeen days after its IPO, on a momentum bet that the AI buildout needs memory more than models. Momentum markets that crown a company this fast can uncrown it just as fast, and ninety days is long enough to separate a durable re-rating from a listing-week spike.

The retail book was 5,526 times oversubscribed at roughly 219 times earnings, on a company whose own prospectus concedes its robots cannot yet do useful work at scale. STAR Market debuts spike in the first sessions when daily limits are off, then reprice as the story meets the first public financials and the lockup-free float. Priced on narrative rather than cash flow, it is the cleanest mean-reversion setup in the current robotics mania, and the closest rhyme is the 2021 pre-revenue EV makers, right about the sector and wrong about the equity.

The Information reports the SoftBank-backed power developer behind OpenAI's Ohio campus aims to list as soon as September, raising at least $5 billion, with Nvidia discussing an investment at the IPO. Power developers are the market's favorite AI adjacency, and this deal tests whether an OpenAI-linked power story can clear the gate while the Ohio financing itself is still being renegotiated.

OpenAI gave Cursor a November 12th shutoff and grounded it in the owner rather than the product: Twitter breaking its OpenAI contract after Musk's takeover, and Musk's own sworn testimony that xAI violated OpenAI's terms. That reasoning does not soften with negotiation, because the thing OpenAI distrusts is the acquirer, which is not changing. Cursor says OpenAI models are only 5% of its traffic, which cuts both ways: it lowers Cursor's urgency to concede and OpenAI's cost of walking. The relationship is also structurally adversarial now, with OpenAI selling a competing coding agent. The main path to the call resolving wrong is a face-saving deal that lets both sides climb down, which is possible but runs against a public statement this pointed.

This is the third reported date change in six weeks: an August target, then September confidential-filing hopes, then October, then November as of this week's WSJ report, each time for a new stated reason. A company that keeps rescheduling around the next quarter's numbers has a pattern, not a plan, and Q3 earnings alone will not settle the safety debate, the antitrust suit, or Nvidia's still-unclosed anchor investment before a roadshow can launch.

Unitree closed its debut day up 629 percent and was trading below half that peak within three weeks; Enflame closed its own debut day at 397 yuan, already short of the 426.54 yuan this ledger's prior call needed. The same aftermarket that stopped rewarding the queue for Unitree has no obvious reason to behave differently for the next chip listing three weeks later.

The lawsuit's entire theory is that the labs' public pacing talk functioned as an enforceable agreement. The best evidence against that theory is what each company does next: Anthropic was reportedly weighing a new model to counter OpenAI's own release the same week it was sued for allegedly agreeing to go slower. A defendant afraid of the suit has every incentive to keep shipping exactly as before, which would prove the agreement was rhetoric, not restraint.

The Lean 4 certificates make outright falsity unlikely. A machine-checked proof is hard to argue with. The live risk here is provenance, not correctness: how the results were produced, not whether they hold. That is what Prediction 2026-08-06-T2 tests.

Underpins both the memory thesis and the memory-bound scaling argument. Capacity takes years to build and the orders are already placed, so the supply side cannot surprise inside the window.

X Square Robot has filed confidentially in Hong Kong, AgiBot has hired three sponsors and begun the process, and Galbot is reported preparing a filing. The AI-datacenter capex cycle and Beijing's embodied-intelligence push have opened a listing window the whole sector is running through at once, and Hong Kong first-half IPO proceeds already ran 92% ahead of last year. With three names in motion, two closing is the base case rather than the bet.

Beijing's algorithm registry cleared Apple's proprietary model in July, the first foreign clearance ever granted, and reporting points to a rollout within months. The regulatory risk is spent; what remains is Apple's execution and its habit of slipping AI dates. If the first foreign-model deployment inside China's wall ships on schedule, the dual-track template becomes the price of admission every other Western company can price; if it slips, the clearance was the easy half.

Dream Security's evidence points to Chinese-language operator documentation; Taiwan's own government has confirmed the intrusion but said only "overseas sources." Formal attribution is a materially higher bar than technical evidence, and it is the fact that would convert a security firm's report into the documented state-linked incident B3 is watching for.

The Journal reports Commerce Secretary Lutnick told Apple the administration opposes its purchases of Chinese memory, after Apple tested CXMT and YMTC chips to ease the shortage. Apple has spent two years courting this administration and defying it over a component swap would be the surprise. The interesting failure mode is the shortage winning anyway: if memory gets scarce enough, even Apple's politics bend, and a teardown would catch it.

The Information reports talks above $30 billion, and the pattern argues for closure: Nvidia has held a stake since 2023, just paid Poolside $6 billion for its model factory, and benefits from anchoring the application layer that markets its silicon. Against it, reported talks die more often than headlines suggest, and Perplexity has enough independent demand to shop terms.

The index printed a record-low 97 cents on August 31st, half its early-summer high, driven by Chinese open-weight releases, OpenAI's July cuts and time-of-day pricing, and Anthropic just cut cache reads 75%. This is the cleanest public gauge of whether frontier pricing power is real or whether inference is commoditizing faster than the capex assumes; a quarter more of decline in four months continues the observed slope.

ARC Prize's Astra evaluation scored 62.7% on its provider-neutral Standard harness and 99.9% on OpenAI's own Provider Adapter harness, a 37-point gap that is itself the finding: a frontier score is also a claim about scaffolding, disclosed or not. Once an independent evaluator shows the gap this publicly, the labs' own single-number reporting becomes conspicuous by comparison.

DeepSeek became the first frontier provider to price tokens by the hour of day, with peak and off-peak tiers and hard concurrency caps that say the move is capacity management. If inference really is capacity-constrained at peak, the rate card spreads the way it did in electricity markets. If it stays one lab's quirk, the utility comparison this morning's editorial leans on is wrong, and we would rather find that out on the record.

The gap between 'results verified' and 'process verified' is the honest criticism nobody is making. Checking a proof tells you it holds; it tells you nothing about how it was found, or what else was tried.

Cadence-based rather than insight-based. The trend line is steep, consistent, and nobody involved has an incentive to stop. The way this fails is a policy shock, not a technical one.

Bloomberg has tallied roughly $70 billion of guarantees, led by Broadcom's backing of chips leased to Anthropic, that sit on no balance sheet any of the guarantors publish. Rating agencies price exactly this kind of contingent liability for a living; the question is how long it takes one of them to say so in writing, which is the mechanism by which F4's credit-cracks-first thesis becomes visible to a reader who does not read bond desks' private notes.

Claude in Chrome went GA with autonomous action: the agent auto-approves steps it judges safe inside a logged-in browser. The browser is the worst injection surface an agent can occupy, because it reads hostile pages as a matter of course, and T4 already bets that no architectural fix arrives this year. A public demonstration is the observable version of that thesis, and the research tooling to find one, ToolHazard included, is now open source.

Anthropic disclosed a model stronger than its flagship and said it has no current plans to release it. Competitive pressure from OpenAI's IPO-season shipping cadence argues for releasing; the Risk Report's own reasoning, a saturated danger benchmark and rising uncertainty, argues against. Which force wins says how much the safety framework actually constrains the roadmap.

They raised guidance by about $1B mid-year, which implies visibility most software companies never have. The risk is timing rather than demand: government contracts slipping a quarter would do it.

All three labs converged on the identical structure this week: full model for everyone, the dangerous tier behind a vetting process. OpenAI's own $1 billion Daybreak pledge is explicitly aimed at exactly this population, and the test of whether gating protects the floor rather than just the labs is whether the floor shows up in the numbers. A subsidy or a vetting program that never discloses who actually got through is indistinguishable, from the outside, from one that reached nobody.

Twenty-five Fields Medal recipients organizing around one complaint, that rushed claims leave no time to check credit, methods, or correctness, is a level of organized pushback this industry has not faced from a scientific field before. Whether it changes anything is a separate, checkable question, and the base rate for public criticism changing a lab's release practice, as opposed to its press strategy, has been poor all year.

Reuters' sourcing puts the figure at up to $10 billion, on top of the Lambda and Nscale cloud commitments Nvidia already backs and the $1.5 billion SB Energy investment from the same playbook two weeks earlier. The pattern this year has been that reported ceiling figures for Nvidia's strategic investments tend to survive negotiation close to the headline number, not shrink toward a token stake.

Google paid $10 million for Spirit Airlines' emails and messages and had to outbid Mercor to get them, which established both a clearing price and a second bidder. Estates owe creditors value maximization, every failed company holds decades of authentic workplace text, and authentic text is the scarcest training input. Once one trustee banks $10 million for an archive previously valued at zero, the next trustee's lawyers will propose it.

The reported structure is already specific: a junior tranche near $30 billion, Blackstone and Apollo at the table, Broadcom guaranteeing part of the senior debt, and a June agreement to fund more than 20 gigawatts of Anthropic compute by 2028 that needs exactly this money. Private credit has more committed capital than eligible AI deals, and every comparable mega-financing this year has closed larger than first reported, not smaller. The call fails if credit spreads widen enough to make the guarantee expensive, or if Anthropic's own IPO changes how it wants to fund compute.

Benchmarks are marketing until silicon carries revenue traffic. OpenAI published performance-per-watt numbers against Blackwell and SemiAnalysis largely endorsed them, but hyperscaler chip programs historically die between the Hot Chips talk and the data center: software maturity, yield, packaging supply and the temptation to keep buying Nvidia all sit in the gap. If the chip is real, production deployment is the test it cannot skip, and OpenAI has every incentive to announce it loudly when it happens.

Nvidia is the dominant AI accelerator maker acquiring the platform that distributes most of the world's open models, including the releases that fill Nvidia's own trending page. Jensen Huang's pledge that Hugging Face 'remains open' and that Nvidia compute will not be required is a promise regulators are positioned to test precisely because the deal is large, cross-border, and touches a chokepoint every lab depends on.

High confidence, because this is an architectural problem rather than a backlog item. Instructions and data share one channel. Until that changes, every defense is a filter. And a filter between a capable search process and its goal becomes part of the search space.

Scale is rentable now. When the compute is a purchase order rather than a moat, incumbency stops holding people, and the reason to stay becomes equity rather than capability.

340% year-on-year growth in attacks, no architectural fix, and agents shipping into production anyway. This is arithmetic rather than prophecy. The uncertainty is disclosure, not occurrence. It will happen; the question is whether anyone says so out loud.

Gartner forecasts over 40% cancellation. That number is unfalsifiable as stated, so this is the same claim rebuilt into a form that can actually be checked.

Secondary-market bid/ask imbalance plus better revenue quality: Anthropic's revenue is more enterprise and more contracted, which is materially easier to underwrite.

The core structural call. Credit is the sensor and equity is the symptom: this build-out began funded by cash flows and is now increasingly funded by debt, and debt reprices before sentiment does.

Tests the Microsoft-concentration thesis from OpenAI's side. Diversifying compute is the tell that the relationship is a supplier arrangement rather than a partnership.

The $500B figure is a target attached to MOUs, and MOU-to-close attrition in infrastructure finance is real. But six institutions with $4T of AUM do not lend their names to a press release they intend to strand, and the demand side (labs converting capex to leases) is visibly forming the same week. One fifth of the headline inside a year is the honest over-under between marketing and a market.

Section 8 of the Clayton Act works by threat rather than litigation: the DOJ's 2022 sweep produced seven director resignations across five companies and zero filed complaints. The statute's remedy is cheap for the target, a resignation, and expensive for the government to litigate, and a16z's political standing raises the cost of filing further. The probe has already run a year without action.

Retaining an adviser is preparation, not an application, and Reuters' sourcing describes people familiar with the matter rather than a company announcement. But DeepSeek has spent this exact week demonstrating it can move fast under pressure, reversing a pricing decision within 48 hours of announcing it, and a lab that has already done the harder work on models and revenue has less to build than a typical STAR Market applicant.

Amodei's essay names a concrete commitment: evaluators with ongoing access comparable to employees, checking safety practices, reporting incidents, and examining training before a model ships. That is a stronger claim than the enrollment-number gap Prediction 2026-09-06-T1 tracks, because it promises an outside voice with something to publish. Anthropic's own week supplies the stakes: a fourth cybersecurity incident, a declined UK safety review, and a resigned researcher all landed before the essay did, so whether an embedded evaluator ever surfaces something publicly is the difference between a governance structure and a talking point.

Six reports in the framework's opening week is a fast start for a voluntary disclosure process, but the framework explicitly separates cases ready for publication from those under extended review, and OpenAI alone decides which cases clear that bar. A process that controls its own throughput and its own definition of what counts tends to slow once the obvious backlog clears, the same pattern Prediction 2026-09-13-T1 is testing for Anthropic's evaluator program.

A company projecting $278 billion of negative free cash flow against $856 billion of infrastructure spending through 2030 cannot fund that gap from operating revenue alone, however fast that revenue grows. Publishing the number itself is the tell: it prepares investors for what comes next rather than describing a plan that closes on its own.

My least confident call, and logged deliberately at 0.60. The interest-rate and over-hiring confound is genuinely hard to separate from an AI effect. At this confidence a miss costs me little and a hit proves little, which is the honest position, not a hedge.

First-generation silicon nearly always slips: tape-out, bring-up, software maturity and datacenter integration each eat quarters, and Fractile has never shipped a chip. The $250 million order is a hedge against Nvidia concentration, not a deployment plan, and Bloomberg itself reports the chips are not expected to be usable until 2027, which leaves no slack for the usual first-silicon respin. If Fractile beats this, it will have done something no first-time inference-chip vendor has done on schedule.

Reuters reports Moonshot negotiating revenue-sharing agreements with all three, seeking up to 30% of K3-related service revenue. The clouds already serve Chinese open-weight models without paying their makers, so a signed deal converts distribution they give away into margin both sides can book. Against it: talks are early, the split mechanics and audit rights are unresolved, and the Treasury Secretary has threatened to blacklist the company, which is the kind of political overhang that kills deals at the signature stage.

A bank is hired and the ask is $13 billion, per Business Insider. Every plausible strategic buyer already sits on the cap table, which lowers diligence friction, and the platform has never been more central to enterprise AI. Against it: exploratory processes run to test a price fail about as often as they close, and the founders have spent a decade choosing independence.

OpenAI paused its largest frontier reinforcement-learning run for two weeks and left it on hold after concluding an internal model (Astra) could reach the Critical cyber tier, and it says the new monitoring adds roughly 20% compute overhead. A lab that has just publicly slowed down, disbanded its standalone Preparedness team, and is walking into an IPO has every reason not to be the first to ship a model it labeled Critical. The call fails the moment capability pressure or a competitor forces the release anyway, which is the interesting thing to watch.

Two deals alone, struck six days apart in the same week Anthropic was trimming Claude Code's own usage limits, already total roughly $80 billion. A company preparing what is reported to be the largest IPO on record is disclosing exactly the scale of forward commitment a prospectus needs to justify a valuation reported above $60 billion in new capital raised, and the number is the cleanest available measure of how much of this cycle's capital promises are real contracts versus press-release aspiration.

The cleanest natural experiment available on frontier pricing power. If a lab can raise the price of its workhorse model and hold it, the race-to-zero story is wrong, and every model of this industry's margins needs redoing.

Anonymous previews on OpenRouter have always resolved to a named lab within weeks, because the free window only pays for itself as a launch. The fingerprinting evidence, tokenizer behavior, refusal patterns and output style, clusters on Chinese labs, with Z.ai's GLM family the leading community theory and Xiaomi's MiMo team second after its CFO teased a new model. A frontier-grade model given away with prompt retention is a data-collection trade that fits a lab preparing an open-weights or aggressive-pricing launch, which is the Chinese playbook this year.

Every major mainland AI-silicon debut this year opened with a pop of this scale, and Unitree closed its first day up 629%. Enflame drew 4,073 times its retail allocation into a 6-billion-yuan deal, so the demand is measured rather than assumed. The interesting failure mode is a fizzle: it would mean the aftermarket has started pricing these businesses rather than the queue for them, three weeks after Unitree's slide from peak began teaching that lesson.

The Information reports the $12.9 billion agreement as done, four days after the sale exploration leaked, and neither company has said a word. If the reporting is right, the companies confirm within weeks, because a deal this size leaks worse unsigned than signed. The risk is choreography: antitrust counsel can hold an announcement past a month-end without the deal being any less real.

September 30th is Moonshot's own stated target, given while denying an August filing, and the pre-IPO round closing today at a reported $50 billion pre-money only makes sense against that timetable. The threat is Washington: a Treasury blacklist threat hanging over the company is exactly the kind of event that slips a filing without killing it.

Nvidia has beaten consistently, and this demand is contracted rather than forecast. Confidence is only moderate because a memory-driven supply miss is a live risk, see F5.

Z.ai says it is holding the weights back roughly two weeks for safety hardening, after reporting GLM-5.3 found 2,436 real vulnerabilities across 269 open-source projects and scored above its own reading of two closed frontier competitors on CyberGym. That is a short, self-imposed window from a lab that has shipped every prior GLM generation openly, and whether two weeks means two weeks is a data point on how a Chinese lab handles a model whose offensive capability outran its own testing.

CNBC's sourcing says the prospectus could go public as soon as the end of this month, and an October listing needs a public filing with road-show lead time. The call tests whether end-of-August sourcing is a schedule or a leak, and it is checkable against EDGAR by anyone.

Grey-market quotes of RMB 410 to 520 imply a 170% to 245% first-day gain and STAR Market debuts have averaged a 466% pop this year, so the call looks safe. It is logged anyway because it pairs with the ninety-day call: a huge debut and a broken offer price within a quarter can both happen, and that pairing is the signature of a momentum market. If even the first day fails, the humanoid pipeline behind Unitree reprices immediately.

Alibaba committed the open weights for its 2.4T flagship to the week of August 10th. Three days in they have not appeared, and the only circulating drop date comes from a single unofficial account. The failure case is the interesting one: it would be the first broken open-weights commitment from a major Chinese lab this year, and it would say more about where Chinese open-weight policy is heading than the release itself.