AGI Has Arrived, Capital Is Losing Its Bearings: Mid-2026 AI Industry & Capital Cycle Review

Kimi K3 (2.8T-parameter MoE, 896 experts with ~104B active, 1M-token context), released July 17, was fully open-sourced late on July 27: weights under a modified MIT license plus three training-infrastructure stacks (MoonEP, FlashKDA, AgentEnv) — handing the community the methodology for training a 3-trillion-class model. Hugging Face CEO Clem Delangue said it topped the trending chart within 30 minutes with 4,000+ likes, the platform's fastest launch ever. On benchmarks, K3 topped the WebDev Arena blind coding leaderboard at 1679 Elo (Claude Fable 5: 1631; GPT-5.6 Sol: 1618) — the first open model to overtake closed flagships in a hard specialty — while ranking third-to-fourth globally on composite intelligence indices (snapshots differ). Day-0 ecosystem support spanned Huawei Ascend 950, Alibaba Cloud, Nebius, Baseten and Fireworks, with Cursor and Cognition (Devin) integrating the model. Competitive responses came fast: Musk reportedly said on July 18 that xAI is training a 2T-parameter model aimed at surpassing K3, and on July 24 Jensen Huang amplified an a16z-initiated open letter, co-signed by Nvidia and others, arguing open and closed frontier models should coexist.

Private-market premia kept expanding: per National Business Daily citing sources, Moonshot AI sent a listing proposal to investors on July 22, targeting an HKEX listing within roughly six months; it is completing a round at a $31.5B pre-money valuation (up more than 7x in half a year, per our prior issue) and plans pre-IPO talks in August targeting $50B pre-money. For DeepSeek, Reuters reported on July 15 a planned round at ~RMB 500B (~$74B) alongside early STAR Market deliberations; yet June's round at ~RMB 450B post-money contrasts with shareholder filings implying ~RMB 350.9B (~$52B) — a >40% dispersion in marks on the same asset, laying bare the fuzziness of private pricing. Zhipu (HK-listed January 8) completed a ~HK$31.4B placement on July 13, extending its research-first narrative. Meanwhile the deceleration turned tangible: on July 28 the Nasdaq 100 fell 10% from its record into correction territory (Bloomberg), the Philadelphia Semiconductor Index logged its longest losing streak of the year, and Nvidia sits more than 15% below its May peak (NYT); Bloomberg attributed the rout to circular-funding worries and Chinese chip competition — underscored by CXMT's Shanghai debut on July 27, which surged as much as 470%.

After confidentially filing its S-1 on June 1, Anthropic — per Bloomberg on July 15 — has Morgan Stanley, Goldman Sachs and JPMorgan scheduling investor meetings ahead of a Nasdaq listing as soon as October; rules require the S-1 to go public at least 15 days before the roadshow. Its May round raised $65B at a $965B post-money valuation. On July 22 AMD announced up to $5B of investment in Anthropic, which will purchase up to 2GW of AMD's latest AI chips from H1 2027 (Reuters) — the invest-for-orders model spreading further. Public markets have begun pricing this private asset: Amazon's July 30 Q2 report showed $62.6B of net income including ~$53.4B of pre-tax, non-operating gains primarily from its Anthropic stake. The control group matters too: OpenAI's listing has slipped toward 2027 (Bloomberg), and SpaceX, which IPO'd in June, now trades more than 40% below its June peak (NYT) — the public anchor cuts both ways, capable of validating premia or puncturing narratives. DeepSeek, per Bloomberg, could file as soon as end-2026 for a 2027 debut.

Verified (Nvidia company statement, August 10): Nvidia, together with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR, is building an independent financing platform targeting more than $500 billion of third-party capital for AI infrastructure. Jensen Huang described compute as becoming an investable asset class; Nvidia closed down 2.86% that day.

Effect on this report's cycle call: this report located top-of-cycle risk at the point where hyperscaler capex growth peaks. If institutional capital can enter compute infrastructure directly through project finance, the cycle gains an additional funding layer — spending is no longer fully bounded by technology-company cash flow and shareholder tolerance. That delays the appearance of the top signal while shifting risk from equity to credit: if compute assets underdeliver on cash flow, the losses land on lenders and the financing vehicle, a transmission path closer to the 2000 telecom fibre cycle than to a simple equity drawdown.

Suggested indicators: track loan-to-value ratios, tenor structures and residual-value assumptions in the first cohort of projects, and watch for the emergence of a rating methodology for GPU assets. These will signal cycle position earlier than absolute capex figures.

The Xiamen keynote hardens this report's headline thesis, "AGI has arrived, capital is losing its bearings," and gives its cycle call, which puts top-of-cycle risk at the point where hyperscaler capex growth peaks, a measurable mechanism. The method for spotting the top is the same: model-layer revenue is an unreliable denominator, so watch only for the moment capex growth starts to slow.

The mechanism is a seven-signal framework. Since March 2026 FutureX has scored valuation, funding pace, new-entrant supply, narrative, fundamentals, liquidity and exits every day. Our August reading: six of the seven are lit. Liquidity has triggered: on 17 August the 30-year Treasury yield touched 5.31% (public market data), crossing our threshold and automatically triggering our accelerated-selling protocol. Exits have only diverged: no AI company listed overseas in Q1, while 155 Chinese companies went public in the first half (public statistics).

Consensus takes three forms, the same three as in 2000. Concentration: the ten largest US stocks make up about 37% of S&P 500 market value, above the 27% peak of 1999–2000 (Goldman Sachs). Circular financing: securitized data-centre debt grew from $1.3 billion in 2022 to about $26.5 billion in 2025, and JPMorgan expects $30–40 billion in 2026 (public statistics). Narrative sameness: model-layer deals fell to 22 in 2025 (36Kr/Zero2IPO); fewer than ten of a hundred companies got funded, an attrition rate above 90%.

Our view: this cycle does not need a verdict on whether a bubble exists. The job is to find the layer where value sits. The bubble is in the middle layer, which carries the most debt and must keep refinancing. The value is in the application layer, which captures what falling costs release, and in the open-source chokepoints every layer passes through.

Full deck: /reports/open-source-breakthrough (first five pages public).

Reported (Reuters, September 4, 2026): Anthropic has pushed publication of its public prospectus to late September, with marketing to begin in mid-October at the earliest and a listing that could land days before the U.S. midterm elections in November; the earlier plan was to publish the prospectus as early as the following week. The company is also finalizing a $15 billion revolving credit facility, and the banks working on the IPO are Morgan Stanley, Goldman Sachs, JPMorgan Chase and Citigroup. Anthropic declined to comment on its IPO plans.

Reported (LatePost, September 2, 2026): Moonshot AI filed a confidential Form A1 with the Hong Kong stock exchange this week, starting its listing process, and is raising a new round at a $50 billion pre-money valuation that may be its last before the IPO. The company closed a Series F of more than $3.5 billion in July at a $35 billion post-money valuation. Moonshot said it does not comment on market rumors.

Confirmed (TechCrunch, September 8, 2026): AI coding company Cognition announced a $2 billion round at a $48 billion valuation, led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst and Avenir; annualized revenue rose from $492 million in May to $900 million, and the valuation nearly doubled from $26 billion in May. Note on figures: Bloomberg reported on September 2 that the round would be about $1 billion at roughly $47 billion; this section uses the company's final disclosure of September 8.

Effect on this report's conclusions: the timetable in the section on exit channels and the Anthropic October IPO needs revising. The public-market pricing anchor moves from October to mid-October or later, possibly close to November. Until then, private-market pricing has no public reference point, and valuations can only be checked against each other through successive funding rounds. The "capital losing its bearings" call has not been falsified in private markets, though its form has changed: money is concentrating further into a few names, and the total has not contracted. Cognition nearly doubling its valuation in four months and Moonshot lifting its valuation target from $35 billion to $50 billion within two months are both evidence of rising concentration.

Reported (Bloomberg, September 13, 2026; first disclosed by Business Insider): Anthropic has selected Nasdaq as its IPO venue. Bankers are discussing a target valuation of roughly $2 trillion, with the roadshow planned for October and a listing targeted before the November U.S. midterm elections; Goldman Sachs, JPMorgan and Morgan Stanley are the lead underwriters. Less than four months after the May Series H at a $965 billion post-money valuation, the pricing anchor has roughly doubled.

Verified (CBS News, September 14, 2026): Over the weekend of September 12–13, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," urging the industry to slow the pace of frontier capability gains; Sam Altman and Elon Musk publicly echoed the call. On Monday, September 14, the Nasdaq Composite fell 147 points (0.6%) to close at 26,186, Nvidia slid 3.4% and Micron 5.3%.

Reported (Bloomberg, September 15, 2026; Cailian Press, September 19): OpenAI is in early, investor-initiated talks on a pre-IPO round at a valuation above $1.2 trillion, up from $852 billion in its March round. An internal presentation that surfaced the same week projects cumulative negative free cash flow of $278 billion over 2026–2030 and roughly $856 billion of cumulative compute and infrastructure spending through 2030; at the current burn rate, existing funds run out in 2028. OpenAI has recently cut token prices to counter Anthropic and Chinese open-weight models.

Reported (Sina Tech, September 25, 2026): Medical AI company OpenEvidence raised $250 million at a $15 billion valuation, led by a16z and several hospital systems, up from $12 billion in January; it has raised more than $1 billion over the past year. Valuations in the vertical application layer keep rising while public markets hesitate on AI spending.

Impact on this report's conclusions: This reinforces Chapter 3 on the exit-channel pricing anchor: Anthropic at about $2 trillion and OpenAI above $1.2 trillion push private valuations into a range public markets have not yet validated, and October's IPO pricing becomes the cycle's first real stress test. It also revises one judgment: the trigger for "capital losing its bearings" is no longer only capital withdrawal but also a call to slow down from the frontier labs themselves, which public markets read as a signal of lower capital expenditure; the September 14 session was the first time that transmission showed up in prices. OpenAI's projected $278 billion five-year cash burn sets a quantitative floor: private markets will need to absorb at least one more round of similar size before 2028.

Reported (draft prospectus seen by Reuters and the Financial Times, relayed by Fortune and The Decoder, September 28-29, 2026): Anthropic's 2025 revenue was about $4.6 billion, roughly 12 times the prior year. Revenue was $4.73 billion in Q1 2026 and $11.5 billion in Q2. The 2025 operating loss was $8.06 billion (versus $2.98 billion a year earlier), and the net loss was about $42 billion, of which about $34 billion came from accounting charges. Two customers made up nearly a quarter of 2025 revenue.

Reported (first by The Information, relayed by Reuters and TechCrunch, September 24-25, 2026): Anthropic's seven co-founders are seeking a combined 50.1% of the vote on most corporate matters through super-voting shares (as long as at least three of them keep a minimum stake), while each holds an economic stake of about 2%. The company is asking shareholders to approve the structure; the Long-Term Benefit Trust would still select most board members.

Reported (Bloomberg, via Yahoo Finance, October 1, 2026): Anthropic is aiming to list as soon as mid-November, could begin marketing its IPO as early as the week of November 9, and is targeting a valuation of up to about $2 trillion. Our September 25 update recorded an October roadshow and a listing before the November 3 midterm elections; the listing date has slipped by about two weeks.

Implications for this report: This strengthens our view that a pricing anchor for the exit channel is taking shape, and it also shows what that anchor costs. Q2 revenue alone is about 2.5 times full-year 2025 revenue, which supports the $2 trillion story. But the 2025 operating loss was about 1.75 times revenue, customer concentration is high, and seven founders holding roughly 14% of the economics would control a majority of votes, leaving public shareholders limited influence over governance. Thresholds: if the valuation is cut below $1.5 trillion after the S-1 becomes public, or the listing slips past the end of November, our "capital losing its bearings" call should move up a notch.