global geopolitics

How DeepSeek and Zhipu AI exposed the assumptions on which $1.8 trillion in Western AI valuations rest

Editorial Analysis | July 2026

I. The Regulatory Hearing as Commercial Theatre

Sam Altman appeared before the Senate Judiciary Subcommittee on Privacy, Technology, and the Law on 16 May 2023, delivering testimony that a senior Democratic senator described as remarkable for being the first instance he could recall of a private-sector executive requesting more government oversight of his own industry. Altman proposed, under oath, a federal agency empowered to license AI models above a defined capability threshold, to mandate pre-deployment testing, and to conduct independent audits of frontier systems. The Washington Post, Time magazine, and Axios all characterised the appearance as a significant and constructive moment in technology governance.

The proposal’s strategic implications received considerably less attention than its theatrical presentation. Licensing regimes in safety-critical industries, pharmaceuticals, aviation, nuclear power, share a common structural consequence: they favour incumbents. A pharmaceutical analogue is instructive. The cost of FDA drug approval, estimated by the Tufts Centre for the Study of Drug Development at over $2.6 billion per approved compound when capital costs are included,1 functions as a barrier that the largest pharmaceutical firms can absorb and that eliminates smaller competitors before they reach the market. A federal AI licensing regime structured around capability thresholds would impose compliance costs calibrated to systems that only the largest, best-capitalised labs currently possess, OpenAI, Anthropic, Google DeepMind, and Meta, while creating legal and bureaucratic barriers that nascent domestic competitors and, crucially, foreign open-weight alternatives could not navigate within the American market.

Altman himself acknowledged that licensing advocacy had strategic dimensions. His 2023 testimony notably omitted any substantive discussion of algorithmic transparency, which the IBM representative at the same hearing raised as a primary concern. In the Questions for the Record subsequently submitted to the Senate committee, OpenAI supported licensing for frontier models while simultaneously defending the retention of model weights and training data within proprietary control. The combination, licensing without transparency, is the regulatory formula most consistent with protecting a closed-source business model from open-weight competition.

By May 2025, the strategic environment had shifted sufficiently that Altman’s second Senate appearance carried a markedly different emphasis. Fortune noted that references to AI safety, which had appeared dozens of times in the 2023 testimony, were largely absent from the 2025 appearance, which was structured around American competitiveness and the speed of the China AI challenge. The pivot from safety to competitiveness as the organising frame of Altman’s congressional engagement tracks with a development in Chinese AI capabilities that the 2023 testimony treated as a manageable background condition but that by mid-2026 had become an acute commercial threat.

II. The Price Competition That Quarterly Reports Could Not Conceal

Coinbase, the cryptocurrency exchange group and one of the larger American enterprise AI consumers, announced in June 2026 that it had redirected routine engineering workloads from OpenAI as the default system to open-weight Chinese models developed by Zhipu AI and DeepSeek. Brian Armstrong, Coinbase’s chief executive, stated publicly that the transition had reduced the company’s AI expenditure by approximately half. The announcement was operationally specific rather than speculative, and it identified a cost differential that independent analysis had been circling for several months.

The pricing comparison underlying that decision is material. Running comparable enterprise workloads through Anthropic’s Claude has been estimated at approximately $4,811 per unit of output, while OpenAI’s GPT-5.5 costs around $3,357 for the same throughput. DeepSeek V4 processes equivalent workloads at approximately $1,071, and Zhipu’s GLM 5.2 brings that figure to roughly $544. The ratio between the most expensive Western system and the cheapest comparable Chinese alternative approaches nine to one. At that cost differential, the question CFOs of large AI-consuming enterprises are asking has shifted from whether open-weight Chinese models are good enough to whether any proprietary American product is nine times better.

The benchmark data suggests the answer is no, by a meaningful margin. Zhipu’s GLM 5.2, which spun out of Tsinghua University’s artificial intelligence programme, achieved a score of 62.1 per cent on SWE-bench Pro, the principal coding performance benchmark used in enterprise AI procurement decisions. OpenAI’s GPT-5.5 scored 58.6 per cent on the same evaluation. GLM 5.2 outperformed the current OpenAI flagship on the benchmark most relevant to software engineering workloads, the single largest commercial AI use case, while costing a ninth of the price.2

DeepSeek’s DSpark product, announced in June 2026, applies a different efficiency methodology: a lightweight draft model proposes outputs that a verification model then checks in batches, producing response acceleration of between 51 and 400 per cent depending on task type. The architectural approach does not require additional training expenditure or chip access above what DeepSeek already commands, it is an inference-stage efficiency gain that compounds atop the existing price advantage. Yuchen Jin of Databricks, the data and AI infrastructure company, characterised DeepSeek as producing the best available output per dollar of inference cost, a judgment that reflects commercial rather than partisan assessment from a San Francisco–based firm whose clients include many of OpenAI’s enterprise accounts.

Anthropic filed a confidential S-1 with the Securities and Exchange Commission on 1 June 2026, targeting a Nasdaq listing at a $965 billion private valuation following its $65 billion Series H financing round led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. OpenAI filed its own prospectus a week later on 8 June at a reported valuation approaching $1 trillion, though Bloomberg subsequently reported that OpenAI’s internal timeline had shifted toward 2027. Both companies’ growth narratives rest on enterprises remaining willing to pay for closed-source Western frontier models at price points that the Chinese open-weight market has systematically undercut. The gap between private market valuation and publicly observable commercial dynamics is sufficiently large that PitchBook analyst Harrison Rolfes identified gross margin, the figure neither company has publicly disclosed in full, as the number that will either validate or collapse the entire narrative when the S-1 becomes public.

III. The Prediction Failure of 2022 and Its Consequences

The October 2022 decision by the Biden administration to restrict Chinese access to advanced semiconductor manufacturing equipment, coordinating through the Netherlands and Japan to limit ASML and Tokyo Electron sales alongside the direct restrictions on American equipment, was accompanied by confident predictions of Chinese AI collapse. Abishur Prakash, co-founder of the Centre for Innovating the Future, told CNBC that month that the export controls had expanded the capability gap to “the point of no return.” Jordan Schneider of the China Talk newsletter characterised the measure as equivalent to annihilating the Chinese semiconductor industry overnight. Both predictions have proved substantially incorrect.

Chinese AI laboratories have responded to hardware constraint by optimising inference efficiency at an architectural level that their American counterparts, with access to abundant NVIDIA H100 and H200 compute, have had less institutional incentive to pursue. The DeepSeek V3 and R1 models were trained on H800 chips, a lower-specification variant that Nvidia produces for the Chinese market within existing export control parameters, at a training cost that the team estimated at approximately $5.6 million, a figure that provoked disbelief among American AI researchers accustomed to training expenditures of hundreds of millions of dollars for comparable capability levels. Whether the $5.6 million figure accurately captures total development cost has been disputed, but the broader point that Chinese AI development has not collapsed under chip constraints is empirically established by the commercial performance of the models now undercutting American market leaders.

The export control regime has produced at least one demonstrable unintended consequence beyond model development. China, restricted from importing high-specification optical interconnect components, has invested in domestic production capacity for the optical networking infrastructure used between AI accelerators and across AI datacentre racks. Chinese manufacturers have since become significant global suppliers in that component category, infrastructure that American AI datacentres now purchase from the suppliers the export controls inadvertently created.

IV. The Semiconductor Market That Western Policy Ignored

The sustained policy and media focus on advanced semiconductor manufacturing, the sub-10nm nodes produced primarily by Taiwan Semiconductor Manufacturing Company and Samsung, and used in AI training and inference accelerators, has drawn attention and capital away from a considerably larger segment of the global chip market. Mature-node semiconductors, typically defined as 28nm to 40nm and above, constitute an estimated 65 to 70 per cent of global semiconductor demand by volume, and they are present in virtually every manufactured electronic product that is not primarily defined by raw computational performance.

Lily Feng, president of Semiconductor Equipment and Materials International (SEMI), told Reuters in an interview earlier in 2026 that approximately 50 per cent of new mature-node fabrication capacity being built globally is located in China, and that China’s share of global mature-node output is projected to reach 42 per cent by 2028, up from 37 per cent in 2026. A 2025 Taipei Times report drawing on TrendForce Corporation data found that by 2027, China’s share of mature-node production is projected to surpass Taiwan’s, while South Korea and the United States, with single-digit shares, are expected to continue declining. The projection comes from Taiwanese industry analysis rather than Chinese government sources, and the institutional incentive runs against overstating Chinese capability.

“Unlike leading-edge semiconductor capacity, which has attracted massive investment and public attention, mature-node production remains constrained, fragmented, and increasingly difficult to expand.” (Doug Getty, Rand Tech, June 2026)

The applications of mature-node chips establish why this market position carries strategic weight independent of the AI narrative. Approximately 80 per cent of the 2,000 or more semiconductors in a modern vehicle are mature-node components. Medical imaging equipment, industrial control systems, telecommunications base stations, defence electronics, and, critically, the power management and networking infrastructure of AI datacentres all depend on chips that the 28nm-to-40nm category encompasses. A Rand Tech report published in June 2026, authored by Doug Getty, warned that mature-node capacity constitutes the semiconductor industry’s largest strategic blind spot, noting that shortages in the category were already observable and that the investment asymmetry favouring advanced nodes had left the dominant chip category structurally underbuilt outside China.

The cost structure of the two segments explains the investment asymmetry. An advanced-node AI chip fabrication plant costs between $15 billion and $30 billion to build, generates substantial media coverage, and attracts political support from governments competing for semiconductor manufacturing prestige. A mature-node fabrication plant costs approximately $3 billion and produces no equivalent political or journalistic interest. The CHIPS and Science Act of 2022, which allocated $52.7 billion to semiconductor manufacturing incentives in the United States, was structured primarily around attracting advanced-node production from TSMC, Intel, and Samsung. Mature-node capacity received comparatively little structural support, consistent with the broader pattern of policy attention tracking the high-status, AI-associated segment of the market rather than the volumetrically dominant and strategically critical one.

Chinese manufacturers, not subject to the same political incentive structure, have pursued mature-node capacity systematically. State-backed SMIC, Hua Hong Semiconductor, and multiple provincial-level investment funds have directed capital toward 28nm production lines throughout the same period that Western governments were competing to attract TSMC’s three-nanometre fabs. The result, as SEMI’s data indicate, is a supply dependency that has been building for years and that the standard Western semiconductor policy debate has consistently underweighted.

V. Open-Weight Models and the Logic of Access Control

The open-source and open-weight AI model ecosystem represents the primary structural alternative to the closed-source licensing model on which OpenAI and Anthropic’s valuations are premised. Meta’s Llama family, released under licences that permit broad commercial and research use, demonstrated that frontier-adjacent capability could be distributed without access restriction. Chinese open-weight models from DeepSeek, Zhipu AI, and Moonshot AI have extended that availability to systems whose benchmark performance approaches or exceeds American closed-source products at a fraction of the inference cost.

The licensing framework Altman proposed before the Senate Judiciary Committee in 2023 would, if enacted, apply restrictions specifically to the capability tier where these competitive dynamics are playing out. A federal agency with authority to license systems above a defined capability threshold, Altman cited the International Atomic Energy Agency as an institutional model, would create a regulatory architecture whose primary practical effect would be to restrict access to open-weight models operating at the frontier, while large incumbent American labs with compliance infrastructure, legal teams, and established government relationships would navigate the licensing process more readily than academic institutions, smaller startups, or overseas developers.

The argument for capability licensing rests on the genuine possibility that frontier AI systems could be misused for serious harm, biosecurity risks, large-scale cyberattacks, and manipulation of critical infrastructure are the cases most frequently cited. These risks are real and deserve policy attention. The problem is that the specific regulatory architecture Altman proposed, centralised licensing, capability thresholds, pre-deployment testing requirements, maps almost exactly onto the architecture that would most benefit OpenAI’s competitive position against open-weight alternatives, whether domestic or foreign. A pharmaceutical executive advocating for FDA approval requirements is not automatically wrong that drug safety matters; the analysis requires examining whether the specific regulatory design proposed matches the stated problem or the proposer’s commercial interests, and in this case the overlap is substantial.

The US government’s AI export control regime, which has placed usage restrictions on five categories of AI products as of mid-2026, adds a further dimension. American frontier AI companies face a compliance environment that their Chinese open-weight competitors, distributing model weights freely under permissive licences to users globally, do not encounter in the same form. A Coinbase engineer switching from OpenAI to GLM 5.2 is making a commercial decision; a Coinbase engineer switching from a US-restricted model to an unrestricted open-weight Chinese alternative is making a decision that the US government’s own regulatory architecture has structured to favour the latter.

VI. What the Valuation Numbers Imply

Anthropic’s Series H valuation of $965 billion, announced in May 2026, and OpenAI’s reported $852 billion valuation following its $122 billion funding round in March 2026 represent combined private-market pricing of approximately $1.8 trillion for two companies that had not yet disclosed their gross margins to public investors. Anthropic’s annualised revenue run rate of $47 billion in May 2026 implies a price-to-ARR multiple of approximately 20 times, while OpenAI’s comparable figure at $852 billion against approximately $25 billion in ARR implies a multiple closer to 34 times. At those multiples, investor return assumptions require either sustained revenue growth at current rates, significant margin expansion as compute costs fall, or both.

The gross margin question is not marginal. Anthropic’s projected operating profit for the second quarter of 2026 was reported at approximately $559 million on $10.9 billion in revenue, a margin of roughly five per cent at the operating level. PitchBook’s Harrison Rolfes stated explicitly that gross margin, not top-line revenue, is the figure that will determine whether the private-market narrative survives public scrutiny. Companies providing AI inference at scale carry substantial compute costs that scale with usage; the economics of the business look considerably different at 40 per cent gross margin than at 60 per cent, and neither company had provided audited public disclosure of that figure as of the IPO filing date.

Chinese competition bears directly on those margin assumptions. If enterprise clients route a growing proportion of routine AI workloads through open-weight Chinese systems at one-ninth the cost, as Coinbase’s decision suggests is commercially rational for standard software engineering tasks, American frontier AI companies face pressure not only on revenue volume but on the composition of the revenue that remains. The workloads that stay in the proprietary Western system will be those genuinely requiring frontier performance, institutional trust, or regulatory compliance — a smaller and more defensible market than the one the current valuations appear to price. OpenAI was reportedly preparing significant token price cuts in mid-2026, a move that would reduce revenue per unit of output while not addressing the structural cost advantage of Chinese alternatives.

VII. Conclusion: The Strategic Picture the Quarterly Reports Are Assembling

The simultaneous emergence of competitive Chinese open-weight AI models and China’s consolidation of mature-node semiconductor manufacturing dominance represents a pattern whose individual components have been widely reported while the aggregate strategic significance has been less consistently articulated. Western semiconductor policy has directed its primary attention and capital toward advanced-node AI chip manufacturing, producing a race to build fewer, more expensive fabs while China has quietly accumulated a dominant position in the chip category that everything, including the AI datacentres those expensive fabs supply, requires to function.

Western AI policy has focused regulatory energy on frontier model capability in ways that structurally advantage incumbent closed-source American companies, while the Chinese open-weight model ecosystem has expanded availability of frontier-adjacent performance to users who have no reason to apply a national loyalty premium to their infrastructure procurement decisions. Coinbase’s CFO, evaluating a nine-to-one cost differential and a benchmark gap that does not favour the more expensive system, did what enterprise CFOs do: chose the option the spreadsheet supported.

The regulatory capture argument does not require attributing bad faith to every safety concern raised by frontier AI executives. Some of the risks they identify are real, and some portion of the regulatory advocacy may reflect genuine safety motivation rather than commercial calculation. What the convergence of incentives, timing, and specific regulatory design proposals does establish is that the two are not easily separated, and that policymakers designing AI governance frameworks should evaluate specific proposals against the structure of the market they would create rather than solely against the stated justification for creating them. An architecture that concentrates frontier AI capability in a small number of licensed American labs while restricting open-weight distribution and disadvantaging foreign alternatives may address some safety concerns; the degree to which it does so relative to the degree it protects the incumbents’ business model is a question the policy process has not yet systematically asked.

Meanwhile, the mature semiconductor question proceeds on its own trajectory, largely independent of the AI governance debate. Chinese mature-node capacity approaching 42 per cent of global output by 2028 is not a projection that can be altered by regulatory frameworks applied to AI models; it reflects capital investment decisions made years ago whose consequences are now becoming structurally embedded. The policy attention that might have interrupted that trajectory was directed elsewhere, toward a three-nanometre fab in Arizona and a public narrative about American semiconductor leadership that the 65-to-70 per cent of the market nobody reported on was quietly qualifying.

Authored By: Global GeoPolitics

If you prefer to make a one time donation in support of my work, you can do so by clicking any link below:

https://buymeacoffee.com/ggtv |

https://ko-fi.com/globalgeopolitics |

Bitcoin: 3NiK8BoRZnkwJSHZSekuXKFizGPopkE7ns

Footnotes

1. DiMasi, Joseph A., Henry G. Grabowski, and Ronald W. Hansen. “Innovation in the Pharmaceutical Industry: New Estimates of R&D Costs.” Journal of Health Economics 47 (2016): 20–33. The $2.6 billion figure is the Tufts CSDD headline estimate inclusive of capital costs and out-of-pocket spending.

2. SWE-bench Pro benchmark results cited in CNBC coverage of Zhipu AI GLM 5.2 launch, June 2026. Zhipu AI is a Beijing-based laboratory spun out of Tsinghua University’s Knowledge Engineering Group.

References

1. Altman, Sam. Written Testimony before the US Senate Committee on the Judiciary (Subcommittee on Privacy, Technology, and the Law). 16 May 2023. https://www.judiciary.senate.gov/imo/media/doc/2023-05-16%20-%20Bio%20&%20Testimony%20-%20Altman.pdf

2. Altman, Sam. Questions for the Record following Senate Judiciary testimony. OpenAI, May 2023. https://openai.com/global-affairs/sam-altman-senate-questions-for-the-record/

3. Armstrong, Brian. Public statement on Coinbase AI cost reduction, June 2026. Reported by CNBC and multiple technology news outlets.

4. CNBC. “Anthropic tops OpenAI as most valuable AI startup, nears $1 trillion valuation in latest round.” 28 May 2026. https://www.cnbc.com/2026/05/28/anthropic-open-ai-startup-value.html

5. CNBC. “The Tech Download: Anthropic’s IPO sets up first big test of AI boom valuations.” 5 June 2026. https://www.cnbc.com/2026/06/05/tech-download-anthropic-ipo-ai-valuations.html

6. DiMasi, Joseph A., Henry G. Grabowski, and Ronald W. Hansen. “Innovation in the Pharmaceutical Industry: New Estimates of R&D Costs.” Journal of Health Economics 47 (2016): 20–33.

7. Feng, Lily. Interview with Reuters on global mature-node semiconductor capacity, 2026. Reported by Reuters and cited by multiple semiconductor industry publications.

8. Fortune. “Sam Altman urges lawmakers against regulations that could ‘slow down’ US in AI race against China.” 8 May 2025. https://fortune.com/2025/05/08/sam-altman-openai-senate-hearing-testimony-china-ai-regulations/

9. Fortune. “OpenAI CEO Sam Altman tells Congress he wants regulation.” 16 May 2023. https://fortune.com/2023/05/16/openai-ceo-sam-altman-congress-regulation-a-i-testimony-eye-on-a-i/

10. Future Digest. “China’s Mature-Node Semiconductor Expansion.” Technology commentary, 2026. Cited in semiconductor industry analysis.

11. Getty, Doug. “Why Mature-Node Capacity May Be the Semiconductor Industry’s Biggest Blind Spot.” Rand Tech, 3 June 2026.

12. Jin, Yuchen (Databricks). Public statement on DeepSeek performance, posted to X, June 2026.

13. Rolfes, Harrison (PitchBook). Analyst statement on Anthropic IPO gross margin significance, cited in CNBC, 5 June 2026.

14. Sacra Research. “Anthropic Revenue, Valuation & Funding.” June 2026. https://sacra.com/c/anthropic/

15. Schneider, Jordan. China Talk newsletter. Commentary on US semiconductor export controls, October 2022. Cited in subsequent analysis of export control predictions.

16. Taipei Times. “China’s semiconductor share to surpass Taiwan’s by 2027.” 2025. Citing TrendForce Corporation data.

17. Time. “OpenAI’s Sam Altman tells senators: regulate us.” May 2023. https://time.com/6280372/sam-altman-chatgpt-regulate-ai/

18. Washington Post. “OpenAI CEO Sam Altman warns of AI’s potential harm, wants regulations.” 17 May 2023. https://www.washingtonpost.com/technology/2023/05/16/sam-altman-open-ai-congress-hearing/

19. Zhipu AI / Z.ai. GLM 5.2 model release and benchmark documentation, June 2026. Company spun out of Tsinghua University, Beijing.

20. US Congress. CHIPS and Science Act of 2022. Public Law 117-167. Signed 9 August 2022.