The global tech sell-off tests the economics behind AI boom

(By Caixin journalists Yang Zirui, Gu Zhaowei and Du Zhihang)

Anthropic’s public stock filing gave battered global technology shares a welcome lift. Whether it can provide investors with something more durable — a reliable valuation anchor for the artificial intelligence boom — is far less certain.

The San Francisco-based model developer published its preliminary prospectus on 26 October, nearly five months after confidentially submitting a draft to the United States Securities and Exchange Commission on 1 June. Its unexpected second-quarter operating turnaround offered rare evidence that building and running frontier AI models can generate operating profits, exceeding consensus expectations.

Relief rippled across global markets: the Philadelphia Semiconductor Index gained 2.82%, Japan’s Nikkei 225 climbed 1.73% and South Korea’s Kospi jumped 3.09%, with the Kospi 200 futures briefly triggering a circuit breaker that halted algorithmic trading.

Yet the rebound followed two bruising selloffs, in July and October, that erased hundreds of billions of dollars in market value across the AI supply chain. 

Nvidia lost more than US$200 billion from its US$5.43 trillion peak; memory giant SK Hynix tumbled over 50% from its high; and specialised cloud provider CoreWeave Inc. shed more than US$35 billion, down more than 55%.

In China, the retrenchment has been steeper. Hong Kong-listed large language model developers Zhipu AI and MiniMax have fallen more than 80% from their peaks, with MiniMax sinking below its offer price. Meanwhile, dozens of humanoid robotics startups lining up for Hong Kong debuts face growing investor scepticism over their commercial viability.

Anthropic’s planned offering — which could target tens of billions of dollars and approach a valuation near US$100 billion — illustrates this precarious balance. While its disclosures could bolster confidence in AI commercialisation, an offering of that scale risks absorbing scarce liquidity from an already strained equity market.

Unlike July’s swift recovery, October’s downturn showed few signs of a floor. In a 14 October report, JPMorgan strategist Nikolaos Panigirtzoglou pointed to a conspicuous absence of dip-buyers, scaled-back corporate share buybacks and historically lean short positions that precluded short-covering rallies.

“The October decline in technology stocks was driven far more by capital flows and sentiment than by deteriorating fundamentals,” Kuang Shi, chief media and internet analyst at GF Securities, told Caixin. Global investors, he noted, are scrutinising the sustainability of capital expenditure, tracking annualised recurring revenue (ARR) at Anthropic and OpenAI, and searching for commercially viable enterprise applications beyond automated coding.

China offers a parallel stress test. Domestic developers are locked in a margin-eroding price war against increasingly accessible US open and closed source models, even as a logjam in Hong Kong’s listing pipeline strains secondary market liquidity. With high-profile developers Moonshot AI and DeepSeek pursuing public listings, their reception will serve as a test for institutional risk appetite.

Across both markets, the debate has changed: the central question is no longer how fast AI demand can expand, but how much capital that growth will consume — and who will absorb the losses if returns disappoint.

Chips under pressure

Nvidia sits at the nexus of the debate, serving simultaneously as the dominant supplier of AI accelerators and as a prolific backer of the companies purchasing them.

Over the past year, Nvidia committed more than US$50 billion across model developers such as Anthropic and OpenAI, neocloud data centre operators including CoreWeave and Nscale and optical networking suppliers Lumentum Holdings Inc. and Coherent Corp.

These arrangements have renewed market scrutiny over “circular financing” —where a vendor provides the finance to customers who direct that capital back into the vendor’s hardware. Critics warn such arrangements risk inflating perceived demand and sustaining unprofitable operators beyond what underlying market economics would support.

On 27 October, Nvidia extended that playbook to Big Tech, announcing a US$3 billion direct investment in newly issued Meta equity. Meta agreed to expand its procurement of Nvidia compute and networking systems while collaborating on workflow optimisation and safety frameworks for autonomous AI agents. Meta shares slipped 1.6% on the news; Nvidia rose 0.9%.

A report from financial data provider the London Stock Exchange Group (LSEG), characterised AI financing not as a closed corporate loop, but as an interconnected network spanning chipmakers, developers, cloud hosts, infrastructure builders and private credit funds. That web, LSEG noted, obscures end-user demand and complicates assessments of credit risk and asset depreciation.

Nomura Securities likened the capital spending race to a classic prisoner’s dilemma: even as free cash flows weaken, tech giants cannot trim capital expenditure (capEx) without surrendering competitive ground. A pushback by shareholders may temper budgets, the brokerage argued, but only a tightening of capital access is likely to halt the cycle.

The semiconductor selloff has spared few. In the two weeks following 6 October, the Philadelphia Semiconductor Index fell roughly 13%, while China’s tech-heavy STAR 50 index dropped 24% from its June high. Nvidia retreated 11%, SK Hynix fell 20% and Arm lost 24%.

Paradoxically, operating results remained robust. Chinese AI chipmaker Cambricon Technologies Corp. Ltd. dropped 13% despite reporting third-quarter revenue growth of 104% and a 148% rise in net profit. GPU designer Moore Threads posted a 1,774% surge in quarterly revenue and swung to a profit, only for its shares to fall 12.5% the following day.

Similarly, Samsung Electronics Co. Ltd. and SK Hynix posted record quarterly earnings in late October and outlined aggressive CapEx expansions for 2027. Yet their shares traded well below June highs.

UBS analysts noted that semiconductor fundamentals have not cracked. Instead, historically elevated expectations left zero room for earnings beats, while a growing camp of institutional investors began pricing in a cyclical peak in AI spending growth rates.

Data centres under strain

Capital strains are most pronounced among “neoclouds” — specialised GPU cloud hosts that have proliferated to meet compute shortages left unaddressed by legacy hyperscalers.

These providers carved out market share through aggressive pricing, rapid procurement and tailored cluster architectures. But the business model demands heavy upfront capital outlays for hardware and physical facilities, while customer billings amortise over multi-year terms. Equipment depreciation and secondary market pricing pressure add to asset-liability mismatches.

CoreWeave illustrates both the boom and the subsequent balance sheet strain.

Founded in 2017 as a cryptocurrency miner, the company pivoted to GPU cloud hosting in 2019, securing early backing from Nvidia in 2023. Revenue surged more than sevenfold to US$1.9 billion in 2024, with Microsoft generating approximately 62% of sales. In the run-up to CoreWeave’s March 2025 IPO, OpenAI signed an US$11.9 billion compute and cloud service agreement and acquired a US$350 million equity stake.

After listing at a US$23 billion valuation, the stock rallied 360% before reversing sharply.

Regulatory filings in August revealed roughly US$51.6 billion in total debt and lease obligations against only US$5 billion in shareholders’ equity. Second quarter operating cash flow of US$679 million fell far short of US$9.4 billion in adjusted capital expenditure, while net interest expense of US$640 million absorbed roughly a quarter of revenue.

Bernstein estimated that fulfilling existing customer commitments will require an additional US$76 billion in outside capital. JPMorgan projected negative free cash flow through 2029, with break-even pushed to 2030.

As rising oil prices rekindled inflation concerns and stoked expectations of tighter Federal Reserve policy in October, yields on CoreWeave’s unsecured notes climbed toward 15%. Its credit-default-swap spreads blew out to roughly 1,100 basis points, and the equity surrendered over 40% in two weeks.

The repricing rippled across the sector. Data-centre operator Cipher Digital experienced a maximum October drawdown approaching 50%; Nebius Group and IREN Ltd. fell roughly 40% from their monthly highs; and Chinese operators VNET Group and GDS Holdings fell by more than 30%.

Physical bottlenecks have compounded financial risks. Power shortages, municipal opposition and regulatory hurdles increasingly delay site commissioning — threatening contracted revenue streams.

On 21 October, Australian operator NEXTDC disclosed that local opposition to grid expansions serving its Melbourne M3 site would delay energisation, prompting it to cut roughly A$140 million (US$98 million) in revenue and adjusted Ebitda from fiscal 2027 guidance. The stock plunged 41.7% in a single session.

In Southeast Asia, Thailand suspended construction across 49 data centres in September and froze approvals for 117 proposed facilities pending national utility and sustainability guidelines.

“Long-term structural demand remains solid, but project delays force a complete recalibration of internal rates of return and debt-servicing schedules,” a private equity executive told Caixin. “Operators facing liquidity pinches may be forced to sell assets at discounts to service interest obligations.”

Model makers confront margin compression

The software developers driving the compute demand face an equally daunting task: converting surging model usage into positive operating cash flows before capital dries up.

Hyperscalers reported formidable top-line momentum in the third quarter. Google Cloud revenue doubled year-on-year to US$30.7 billion; Amazon Web Services rose 39.3% to US$46 billion; and Microsoft Azure grew 48% to US$33.2 billion.

Yet capital expenditure surged alongside revenue. Alphabet spent US$51.4 billion in the quarter, tipping free cash flow into negative territory. Amazon spent US$77.3 billion — up 120% and well ahead of forecasts. Microsoft’s annual capex reached nearly US$116 billion for the fiscal year ended 30 June, an 80% year-on-year increase.

Meta’s standalone AI assistant, Muse, briefly reached No. 1 on Apple’s US App Store downloads following its September launch. But consumer enthusiasm did little to assuage investor unease over hosting costs. Meta posted negative third-quarter free cash flow of US$2.12 billion and unveiled a US$35 billion equity offering on 26 October.

GF Securities’ Kuang noted that while Muse demonstrated technical progress in autonomous agents, Meta must still prove the product can perform complex tasks reliably at manageable inference costs.

In China, listed developers Zhipu AI and MiniMax have suffered an even sharper valuation derating.

Both posted strong top-line growth in the first half: Zhipu’s revenue climbed roughly 400% to 954 million RMB (US$143 million), while MiniMax expanded 283% to US$117 million. Although both narrowed their net operating losses, both remained deeply cash-flow negative.

By 27 October, Zhipu was down 82% from its June peak. MiniMax fell below its HK$165 (US$21) IPO price on 22 October, off 88% from its March high.

At their peaks, both traded at multiples exceeding 250 times ARR. Following the rout, Zhipu’s multiple contracted to roughly 21 times Annualised Recurring Revenue (ARR), and MiniMax’s fell to eight times.

By comparison, Anthropic’s US$69 billion in July ARR and a targeted private valuation between US$1.5 trillion and US$2 trillion imply a forward multiple of 22 to 29 times.

“The drop reflects a necessary valuation correction rather than a broken business model,” a technology analyst at a Chinese brokerage told Caixin. “However, aggressive price-cutting by global developers has compressed pricing power across the domestic tier.”

Margin erosion is already evident. Zhipu’s first-half gross margin fell to 26.4% from 50% a year earlier; MiniMax registered a gross margin of 17.9%. Both completed dilutive capital raises over the summer to fund ongoing training and inference clusters.

China’s IPO chill

The tech repricing has dampened China’s primary listing markets.

Hong Kong hosted 116 public debuts in the first nine months, up 76% year-over-year, raising HK$387.9 billion. But secondary performance turned negative in September, with four out of five mid-month debutants breaking below their offer prices on day one, according to Li Feng, founding partner of Frees Fund.

That marked a sharp shift from the first half, when only 14.3% of listings broke issue prices — the lowest break rate in three years.

A growing supply of new shares has collided with slowing portfolio inflows. More than 380 companies were awaiting listing hearings in Hong Kong as of 28 October, according to market tracker iFinD, while foreign funds logged net outflows of HK$15.9 billion during the first trading week of October, according to Guosen Securities.

“The structural constraint in Hong Kong is market absorption capacity,” an equity strategist told Caixin. “The supply of new paper is outpacing the pool of secondary capital available to support it.”

Moonshot AI, the startup behind the Kimi chatbot, has confidentially submitted an application for a Hong Kong IPO and informed investors it is awaiting regulatory approval, Caixin learned.

An executive at a Chinese private equity firm said the timing and valuation of Moonshot’s listing will be a critical gauge for the market. Rival DeepSeek is concluding a pre-IPO round and preparing an application for Shanghai’s STAR Market.

For embodied AI and humanoid robotics, the market pullback has been swift. During July and August, between 30 and 50 robotics startups were preparing Hong Kong listings, often on early-stage commercial narratives.

By September, institutional appetite had waned. Leju Robotics Technology Co. Ltd. suspended its Shenzhen IPO review on 30 September; Zhejiang Fine Motion Robot Joint Technology Co. Ltd. withdrew its Shanghai application on 21 September. Market participants report that several planned Hong Kong listings for embodied AI developers have been put on hold.

Revenue quality has also drawn scepticism. Several robotics startups relied heavily on bulk equipment sales to state-backed “data-collection testbeds,” raising doubts over commercial repeat orders.

According to venture database IT Juzi, China’s robotics sector completed 19 financing deals totalling 4.93 billion RMB (US$740 million) in the first half of October, down from 91 transactions totalling 31.32 billion RMB across August.

Repricing the robotics stack

Even as private funding slows, shifts in foundation models are upending the robotics value chain.

On 22 October, Unitree Robotics founder Wang Xingxing demonstrated OpenAI’s GPT-6 Astra controlling a bipedal robot without task-specific tuning. In the demonstration, the robot completed a multi-step watercolour painting task, adjusting its grip and rinsing brushes.

Wang projected that general-purpose household robots could enter out-of-the-box consumer deployment by 2027. Unitree shares, which had dropped post-debut, rebounded over 71% between 16 and 28 October.

Yet the emergence of generalised foundation models capable of direct physical control threatens startups whose core product is a proprietary, standalone robotic “brain.”

“If a humanoid-robot company isn’t full-stack across both hardware and software, and only focuses on developing the robot’s ‘brain,’ it is in a perilous position,” a technician at a robotics company told Caixin.

For investors, that creates another valuation problem: advances in general-purpose AI may expand the broader market for robots, but they risk rendering obsolete the proprietary models of startups funded solely to build specialised robot brains.

Han Yuhang, Qu Yunxu and Ge Haoyue contributed to the story.

This article was first published by Caixin Global as “Cover Story: The Global Tech Sell-Off Tests the Economics Behind AI Boom”. Caixin Global is one of the most respected sources for macroeconomic, financial and business news and information about China.

Related: Inside the US$50 billion rise of Moonshot AI | AI drives markets as valuations race ahead of earnings