The evidence, which dates mostly from mid to late October 2026, supports a distinction that is easy to blur. In the short run, Nvidia's position looks secure, because demand is described as surging 8,10 and rival silicon has not yet displaced it. In the long run, the structure around Nvidia is changing in three ways. Its largest customers are building alternatives. China is constructing a partly separate stack. Much of the financing that sustains purchases is increasingly circular, and Nvidia itself helps to provide it. The interesting question is not whether Nvidia's demand is large, but how much of it rests on capital that Nvidia and its customers are supplying to each other, and how quickly the buyers can adjust away from it.
Many individual claims below rest on a single source. I say so where it matters.
Nvidia as supplier and financier
The Meta and OpenAI cases
Nvidia is reported to be investing $3 billion in Meta 73,75. Meta, its largest single customer 5,80, has announced that it wants to raise up to $95 billion 34,60. A later summary says the plan was expanded to $97.5 billion 72,82. Meta says the proceeds are for AI infrastructure 55,56. This sits on a February multi-year partnership covering millions of GPUs, CPUs and networking technology 55. Capital thus flows to a large buyer that is itself raising money to buy more hardware. That reading is my interpretation, but the underlying facts agree across several sources.
The OpenAI relationship shows the same mechanism with more risk attached. Nvidia holds an equity stake in OpenAI 1,3,35,77. For the proposed Pike County, Ohio data center, it can take over the lease if OpenAI misses payments or becomes insolvent 77,88. The guarantee is described as up to $105 billion 88. It ends if OpenAI posts its own collateral 77 or reaches a sufficiently high credit rating 77. A backstop of this kind turns a supplier into a contingent counterparty, so Nvidia bears part of its customer's credit risk. Nvidia is also described as working on financing platforms with Apollo 29 and as having partnered with BlackRock on compute financing in August 48.
The support is selective. One report notes that the largest recent commitments by Meta and AWS had no disclosed Nvidia financing 39. Nvidia's financing is therefore a tool applied where it is needed, not a general feature of every sale.
The circularity critique
Outside commentary calls this circularity. The IMF is reported to have tracked 14 public transactions related to circular AI investment since 2023 and found $300 billion invested in financially weaker AI companies 90. The Financial Times is cited as saying that $160 billion of $285 billion in Q2 2026 profits at Alphabet, Amazon, Meta, Microsoft and Nvidia came from paper gains on stakes in other AI companies 78. Each of these figures has one source in this material. They do fit the specific deals above, and they suggest that part of reported prosperity depends on valuations holding.
The demand chain behind the orders
Concentration and its financing
Microsoft, Meta, Amazon and Alphabet are said to account for roughly half of Nvidia's data center sales 45. Their ability to keep buying depends on how the spending is funded. Combined capital expenditure at Amazon, Microsoft, Alphabet, Meta and Oracle is reported to exceed their combined operating cash flow for the first time 86. The figures supplied, about $416 billion invested in 2025 86 and about $707 billion of operating cash flow 84,86, do not themselves show that crossover. The material does not reconcile them, so the claim is best read as directional. Goldman Sachs is cited as saying Big Tech will fund a third of AI investment with debt 13. Meta raised $24.91 billion in bonds in Q2 2026 71. SpaceX closed about $40 billion of GPU financing with Apollo funds, which buy the chips and lease them back 43,46.
Revenue coverage and the two-lab dependency
The weak point is revenue. One analysis puts Microsoft's AI revenue-to-capex coverage at about $0.19 per dollar and says the industry ratio is similar 89. Anthropic and OpenAI are estimated to account for more than 80% of the expected $165–220 billion end-2026 AI model and app run-rate revenue 58. Reported figures for the larger of the two have been unstable. OpenAI's run-rate fell from $70 billion to about $50 billion after a Financial Times correction, because the earlier figure counted gross transaction value 13. OpenAI also raised its compute spending forecast through 2030 to roughly $750 billion 47. Commenters argue that the failure of either lab, or disappointing demand, would reprice AI capex broadly 12. Others dispute whether other customers could replace them 12. That dispute is the crux, because it concerns the elasticity of the rest of the market in absorbing capacity that the two labs might leave unused.
The counterargument
The supportive evidence deserves equal weight. Meta says its AI investments are already generating returns through better advertising performance 79. Hyperscalers are reported to view under-investing as more catastrophic than over-building 89. Microsoft's AI business has reached a run rate above $37 billion, a figure reported by six sources 2,7,9,30. Even so, Microsoft Cloud gross margin fell to 65%, which the post attributes to AI infrastructure investment pressure and service mix 81. Demand is real, and its financing is the open question.
The edges show some strain. Microsoft withdrew as anchor tenant of Nscale's Loughton, Essex project, which was to provide 50 MW for Azure 15,36. Microsoft says it remains committed to the partnership 36, and the episode is attributed to power delays 20. A "neocloud" shakeout is also being discussed 22. These are isolated reports, but neoclouds are one channel through which Nvidia hardware is financed and deployed, so the pattern is worth watching.
Substitution: gradual, and aimed at the software layer
Hyperscalers and OpenAI
Amazon, Google and Microsoft are all described as developing their own chips 57. Microsoft unveiled Maia 300, its third-generation custom chip 24, with at-scale use reported for 2027 23. The stated aim is to reduce reliance on Nvidia 24. Two qualifications keep this in proportion. One post says efficiency claims for the Maia accelerators remain an open question 69. Another framing says Big Tech is going custom "but Nvidia still leads" 33.
The more instructive development concerns software, where lock-in is formed. Google open-sourced TorchTPU 44. Amazon responded with TorchNeuron, which it says lets developers deploy PyTorch workloads on AWS without massive retooling 44. Amazon is also contributing upstream code to decouple PyTorch components from Nvidia GPU assumptions 44. The article concludes that ecosystem lock-in will face more scrutiny 44. This approach reduces the switching cost for each buyer, which is the friction that sustains a premium, and that matters more than any single competing chip.
OpenAI is diversifying on several fronts. It developed the Jalapeño accelerator with Broadcom 4,74. It has a 6 GW commitment with AMD 37,38 and collaborates with Broadcom on accelerators covering 10 gigawatts 6,40. A widely repeated claim says OpenAI plans to buy Google AI chips. It circulates as a post attributed to The Information 76, but the material stresses that it is unverified, rests on a single short post, and has no OpenAI confirmation 76. I treat it as a signal of sentiment, not a confirmed shift. OpenAI is also bringing Nvidia Vera CPUs into production on Oracle Cloud Infrastructure 83, so its Nvidia ties persist alongside its diversification.
Partner, customer and substitute at once
Microsoft shows that "competitor" is too crude a label. It is a launch partner for RTX Spark 41. It co-launched the Physical AI Growth Lab with Nvidia, Infineon and UnternehmerTUM 17, and its Azure platform supports Nvidia's OpenShell Reference Architecture 63. It is simultaneously a major customer, a partner and an emerging substitute. Custom silicon elsewhere points the same way. ByteDance is reported close to buying custom chips from MediaTek 49, which would make it MediaTek's second major ASIC customer after Google 32. Most such claims have a single source, so they show direction, not measured share loss.
China: partial substitution, partial leakage
China offers the clearest case of substitution in progress, and the best-corroborated data point is Huawei's share of China's AI accelerator market. IDC figures relayed in several posts put it at 28% in the first half of 2026 25,26,27,67, within total shipments of 2.008 million units 25,26. One article says Chinese AI-chip investment was redirected from H100/H200 procurement toward domestic accelerators 51. DeepSeek is planning to deploy 160,000 Huawei AI chips 31, and its DeepInfer inference framework lists the Huawei Ascend 950 NPU as supported hardware 42. Meituan says it trained a 1.6-trillion-parameter model entirely on domestic chips 61.
Substitution in the short run is limited by capacity. Huawei and other domestic suppliers cannot meet all demand for AI chips 49, and restricted hardware still arrives indirectly. Chinese AI labs are said to reach compute through overseas cloud providers and third-party resellers 52. One IFP-based estimate puts China's offshore access at 1 million H100-equivalents 70. A Super Micro-linked $2.5 billion server shipment was routed through Southeast Asia to mainland China 59. Taiwan plans to tighten export controls on AI servers bound for China 50, but the scope definition has not been published 62. China is moving toward Huawei while remaining partly dependent on Nvidia-class silicon obtained indirectly.
Efficiency gains that still run on Nvidia hardware
The technical evidence shows how this competition cuts both ways. On the Artificial Analysis AA-AgentPerf benchmark, DeepSeek V4.1 Flash delivers about 1.3x the throughput per megawatt of V4 Pro 0813 on GB300 at 100 tokens/s per user. This is the most heavily corroborated technical claim in the material, with six sources 66. The model is reported to run on as few as 2 GPUs 66. Nvidia reports a 2x speedup in one month on DeepSeek V4 Pro using AI-assisted engineering 68. Jensen Huang has publicly defended Nvidia's use of Chinese open-source models 14. The inference I draw is that cheaper Chinese models are being made more efficient on Nvidia systems, so Nvidia partly benefits from the commoditisation of models. The benchmark does not state that conclusion itself.
One discrepancy remains. Sources disagree on the model's active parameters: 8B per input token in one report with two sources 66, 16B in another 66. The lower figure has more support, but the disagreement is unresolved.
Broadening the base as a hedge
Nvidia's own conduct is consistent with an effort to reduce dependence on a few buyers. It has signed a partnership with the German government on industrial and physical AI 18, with media reporting that several hundred million euros were intended 86. Huang committed $250 million in compute over five years for British AI safety research 45 and pledged an additional $2 billion in European AI deals 45. Other efforts include automotive partnerships such as Mercedes-Benz, reported by five sources 53,54,64,65, and an IREN partnership supporting up to 5 gigawatts of potential deployment 85. Huang has also cited early infrastructure build-outs in Africa, Central Asia, Southeast Asia and Latin America 87.
These programmes should be read as long-run adjustments, not short-run offsets. The material gives no figures showing that they are yet large relative to hyperscaler demand, and most lack disclosed financial terms.
What the evidence implies
Three conclusions follow, each conditional on the evidence available.
- Nvidia's growth is tied to its customers' access to capital. The Meta investment, the OpenAI stake and lease backstop, and the Apollo and BlackRock work all point this way. Customer financing exposes Nvidia to credit risk as well as demand risk.
- Substitution is advancing, but it is gradual and uneven. Maia 300, TorchTPU, TorchNeuron, Jalapeño and Huawei's 28% share are real, but Maia's efficiency is untested and the OpenAI–Google chip purchase is unverified. China's substitution coexists with substantial indirect access to restricted chips. The software-layer efforts are the most consequential, because they lower switching costs.
- The decisive variable is the demand chain, not any single rival chip. Capex is reported to exceed operating cash flow among the major buyers, AI revenue coverage of capex is low, and two labs dominate AI revenue. Equity sentiment reflects some unease: one tracker counts 9 of the 12 most-searched AI stocks down 10% or more from their 3-month highs 11,21, though other snapshots give 8 16 and 7 19,28. The count differs by date, but every version shows most names well off their highs.
The sources do not show that demand itself has weakened. They show a market whose financing is stretched and whose largest buyers are slowly widening their choices. The evidence that would change this view is concrete. It would be OpenAI and Anthropic revenue holding against their compute commitments, disclosed financial terms for the sovereign and enterprise programmes, and confirmation of whether Maia and the PyTorch-based alternatives achieve efficiency parity at scale.