Nvidia & TSMC — AI Compute Infrastructure

Nvidia designs the GPUs that train and run frontier AI models, and TSMC (Taiwan Semiconductor Manufacturing Company) fabricates nearly every leading AI chip, including Nvidia's. Together the two companies form the dominant hardware layer of the AI stack, a concentration that bears directly on compute governance, export controls, and AI sovereignty. The Stanford HAI AI Index 2026 describes the reliance on a single Taiwan foundry as "a single-point-of-failure vulnerability" (Stanford HAI AI Index Report 2026).

Nvidia

Overview and role

Nvidia designs the dominant GPUs for AI training and inference. METR time horizons, scaling laws, and frontier model training all depend on Nvidia hardware. The company's revenue is driven by AI data-center demand; it has continued to post record revenues, including in the modeled 2028 Global Intelligence Crisis scenario where the broader economy showed stress. Nvidia is one of the largest investors in Anthropic.

Financials and stock

Nvidia reported Q2 FY2027 earnings on August 26, 2026: revenue of $96.2 billion with data-center revenue of $89 billion, up 117% year over year, and third-quarter guidance of $108 billion. Committed spending to secure supply and manufacturing capacity rose to $279 billion from $119 billion the prior quarter (Source: techcrunch.com).

On the same day's call the company issued its first year-ahead revenue forecast, projecting roughly 70% growth for fiscal 2028 against analyst models of about $570 billion, or 44% growth. Jensen Huang framed the number as supply-bound rather than demand-bound: "Even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%." Colette Kress guided third-quarter gross margin to 74%, down from 75% in the second quarter, and said margins would bottom at 71–72% in the fourth quarter before settling around 72–73% next fiscal year as price increases take effect. Addressing the circular-financing criticism recorded below, Kress said: "We recognize the scale of this support, and we know some will call this circular financing. We see it differently." Nvidia disclosed maximum gross guarantee exposure of $108.5 billion, most of it credit support for SB Energy's Ohio campus hosting Nvidia compute leased to OpenAI. The stock rose more than 4% in after-hours trading (Source: fortune.com; sec.gov).

Nvidia reported Q1 FY2027 earnings on May 20, 2026, for the quarter ended April 26, 2026.

Nvidia's private-company portfolio, valued at $43 billion, nearly doubled in a quarter where its hardware revenue grew 85%, revenue partly fueled by sales to companies in which Nvidia is invested (OpenAI, Anthropic, and others) in vendor-financing-style structures. This holding is cited as central evidence for the Circular Financing in AI thesis. (Source: techcrunch.com) See Circular Financing in AI, AI Bubble Debate.

Snapshot — Stock and market cap

By July 8, 2026, Nvidia's market value had slid by roughly $1 trillion from its April peak, sending its valuation multiple back to pre-AI-boom levels, per Bloomberg (Source: bloomberg.com). The slide extended the June 5 chip selloff described below. See AI Bubble Debate.

On June 5, 2026, U.S.-traded chipmakers shed roughly $1.3 trillion in market value as the PHLX semiconductor index fell 10.3%, its worst single day since March 2020. Nvidia fell about 6% (shedding more than $300 billion), with Micron down 13% and Marvell down 17%, extending losses triggered by Broadcom's weak custom-AI-chip outlook disclosed two days earlier (a June 4 revenue miss). (Source: reuters.com) The selloff co-occurred with NSPM-11, the OpenAI government-stake talks, and the SpaceX–Google compute lease. See AI Bubble Debate, AI Bubble vs. Buildout — Synthesis, AMD — Advanced Micro Devices.

Treatment in US national accounts

An August 2026 Epoch AI report argues that US GDP statistics miss most of Nvidia's value-add because of its fabless supply chain. Nvidia sends designs to foreign contractors such as TSMC, sells finished chips to foreign server makers, and those server makers sell servers to US data-center owners; the servers enter the US at a price that includes Nvidia's margin and are subtracted as imports, but no matching export is recorded, since the chips never cross the US border and no foreign buyer pays explicitly for Nvidia's IP. The report states it confirmed this with the Bureau of Economic Analysis. It reports that Nvidia made around $120 billion in US pre-tax income between January 2025 and January 2026, attributes to Nvidia a measurement gap averaging about $30 billion per quarter through 2025, and, using Nvidia's US-booked operating income as a proxy, estimates that annualized US GDP growth over the prior year was underestimated by about 0.3 percentage points, potentially almost two percentage points per year by 2028 if Nvidia's operating-income growth continues. Extending the analysis to Qualcomm, AMD, Marvell and Broadcom would barely change the estimate, according to the report (The Nvidia-sized hole in US GDP statistics (Epoch AI, August 2026)).

Acquisition of Hugging Face

Nvidia agreed to buy Hugging Face for $12.9 billion, The Information reported on August 26, 2026, citing a person with knowledge of the agreement; neither company responded to requests for comment. Nvidia had been among the backers, with Salesforce and Alphabet, of a $235 million round in 2023 that valued Hugging Face at $4.5 billion, and Hugging Face had rejected a $500 million Nvidia investment offer valuing it at $7 billion, per Financial Times reporting in January 2026. The rationale attributed to Nvidia's leadership is consistent with its stated position on open-weight models recorded below: that a plentiful supply of successful open models preserves the company's hardware position by acting as a counterweight to closed-source developers building competing server chips (Source: theinformation.com; reuters.com).

Nvidia announced a definitive agreement on September 3, 2026 at $12.93 billion, of which about $11.9 billion goes to Hugging Face's investors, alongside an equity-based retention program of up to $1 billion for employees who join Nvidia. Jensen Huang wrote that Hugging Face "will remain an open platform for the entire AI ecosystem" and that Nvidia systems would not be required to build on or deploy through it, noting that Nvidia has released more than 500 models and more than 250 open datasets there. Delangue told CNBC that Hugging Face approached Nvidia a few weeks earlier while holding other bidders (Source: axios.com; variety.com; calcalistech.com). Huang framed the purchase in his September 3 post as strengthening safety, cybersecurity and sovereignty rather than only as a distribution acquisition, a framing amplified by Trump administration figures and by advocates of accelerated AI development (Source: insideaipolicy.com). Hugging Face was last valued at $4.5 billion in an August 2023 round that raised $235 million, so the agreed price is roughly 2.9 times that mark (Source: reuters.com).

Amazon partnership expansion (August 2026)

Nvidia and Amazon announced an expanded partnership on August 26, 2026 adding 2 million Nvidia GPUs — Blackwell Ultra, Rubin, and Rubin Ultra — to Amazon Web Services data centers across 2027 and 2028, five months after Amazon agreed to deploy more than 1 million Nvidia GPUs across AWS. Neither company disclosed financial terms. Colette Kress said an unspecified number of Vera CPUs would accompany them, and that Amazon would adopt Nvidia's Omniverse, Cosmos, Isaac, and Jetson stack for its robot fleet (Source: techcrunch.com).

GPU backstop and revenue-share financing

Nvidia said in reporting dated July 1–2, 2026 that it would take a cut of some customers' cloud revenues in exchange for using its balance sheet to help those companies finance purchases of its AI chips (Source: theinformation.com). Under the program as detailed by Data Center Dynamics, Nvidia financially backstops neocloud customers' GPU purchases by agreeing to rent back unused GPUs at a fixed rate in exchange for a share of cloud revenue; Firmus (deploying 170,000 GPUs in Batam, Indonesia) and Sharon AI (40,000 GB300 GPUs) are the first adopters, following 2025 backstop deals with CoreWeave ($6.3 billion) and Lambda ($1.5 billion) (Source: datacenterdynamics.com). The arrangement extends the vendor-financing posture documented in Circular Financing in AI.

The largest reported extension of the posture came on July 26, 2026, when Nvidia entered talks with OpenAI to guarantee roughly $250 billion in financing covering the lease and debt on a 10-gigawatt data center that SoftBank's energy subsidiary is developing in southern Ohio. The project is expected to cost more than $500 billion including chips, and Nvidia was separately reported to be discussing financing up to $350 billion of OpenAI chip purchases. The guarantee would be OpenAI's first step toward controlling infrastructure rather than renting it from Microsoft, Amazon and Oracle; negotiations were characterized as early and capable of collapsing (Source: bloomberg.com; wsj.com).

The structure was reworked over the following weeks to reduce Nvidia's exposure. People familiar with the deal said on August 14, 2026 that Nvidia would initially guarantee only half the 10-gigawatt build-out, lowering the guarantee from $250 billion to less than $120 billion, a change made to address investor concern about Nvidia's risk; no deal had been signed. Four people with knowledge of the talks put the figure at around $100 billion in credit support covering roughly the first two-year phase, about half the total project (Source: wsj.com; theinformation.com). Reporting on August 17, 2026 described Nvidia as close to an agreement on the roughly $100 billion figure, still applying to the first phase of the project, and noted the figure remains lower than the earlier Wall Street Journal report of a possible $250 billion backstop (Source: theinformation.com). The transaction was signed later that day: Nvidia agreed on August 17, 2026 to spend as much as $105 billion backing the first phase of the Pike County complex that OpenAI will lease from SoftBank's SB Energy, with OpenAI securing up to roughly 8 gigawatts of computing capacity and the first 800 megawatts expected online by 2028 (Source: bloomberg.com). The Wall Street Journal described the same agreement as a 10-gigawatt deal with SB Energy partly backed by Nvidia, structured to help the developer raise debt financing without exposing the chipmaker to substantial risk; the two capacity figures differ and neither article reconciles them (Source: wsj.com). Axios characterised Nvidia's commitment differently, reporting that Nvidia agreed in an SEC filing to guarantee up to $105 billion in conditional lease and power payment obligations to SB Energy rather than to spend that amount itself, that Nvidia is the campus's exclusive chip supplier under a 20-year lease OpenAI signed for the site, identified as PORTS-Pike, and that Nvidia is separately investing $1.5 billion in the developer. Axios described the site as an 8 IT-gigawatt facility powered by 10 gigawatts of new generation, with initial capacity scheduled for 2028 (Source: axios.com). Jensen Huang rejected characterisations of the structure as circular financing, writing on X that "OpenAI will pay the lease" (Source: axios.com). See Circular Financing in AI.

Nvidia and Safe Superintelligence announced a long-term strategic partnership on July 27, 2026, giving Ilya Sutskever's lab access to the Vera Rubin platform and roughly tenfold compute growth, shifting it off its prior reliance on Google TPUs. Nvidia described its investment only as "substantial"; Reuters and Bloomberg put the equity investment at $5 billion, citing people briefed on the deal (Source: wsj.com; reuters.com; siliconrepublic.com).

Bloomberg reporting published July 27, 2026 put the disclosed deal pipeline above $750 billion, against more than $540 billion of similar deals Nvidia has announced in 2026 to date. The July 24 SK Group partnership covers more than $500 billion in two-way business and more than 2 gigawatts of AI data centers on the Korean Peninsula, the first built by SK Telecom and opening in 2027. On the same day the pipeline figure was reported, Nvidia shares fell 5% to $196.51, the worst single-day drop since June 5, and the company fell below Apple as the world's most valuable listed company; the cost of protecting its debt against default rose by as much as 0.14 percentage point, to 0.82 percentage point a year. Jensen Huang has called the circularity characterization "ridiculous" (Source: themalaysianreserve.com). See Circular Financing in AI.

Nvidia said in a statement on August 10, 2026 that Apollo Global Management, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs and KKR would partner with it to source $500 billion in financing for AI infrastructure, creating "dedicated pools of capital at significant scale at attractive rates for Nvidia customers." Jensen Huang said in a CNBC interview that he approached only those six firms and that none turned him down, and that "this is really the first time that technology chips have become an investable asset class"; BlackRock chief executive Larry Fink called the effort the start of the "next future for financial engineering," likening it to the creation of mortgage-backed securities in the 1970s. The capital, which Nvidia describes as entirely third-party, is intended to flow to Nvidia's customers — hyperscalers, frontier AI labs and enterprises — rather than to Nvidia itself (Source: cnbc.com; nvidianews.nvidia.com).

Two qualifications surfaced the following day. The Information reported on August 11, 2026, citing a person close to the agreement, that the arrangements are memorandums of understanding rather than completed agreements, and that Nvidia holds an option to backstop up to 25% of the financing for any project in the partnership (Source: theinformation.com). The backstop option places the announcement within the vendor-financing posture described above rather than outside it, since the third-party characterisation holds only to the extent the option goes unexercised.

The Wall Street Journal's account of the same August 10, 2026 announcement described the arrangements as preliminary agreements to establish compute financing platforms, under which special-purpose entities would issue debt and lease the hardware to Nvidia customers with the compute itself serving as collateral. It characterised the more than $500 billion as an uncommitted target over an unstated period, noted the agreements remain subject to execution of final documents, and reported that no interest rate, maturity, deployment timetable or first project had been published (Source: wsj.com). Stijn Van Nieuwerburgh's Brookings Papers on Economic Activity paper (Fall 2026) describes the arrangement more firmly, as a $500 billion credit facility "stood up" in August 2026 for which NVIDIA "will provide a 25% residual value guarantee, exposing it to as much as $125 billion of backstops." He cites it as an example of the circularity created when a dominant equipment supplier helps finance customers' purchases of infrastructure built around its own products, and adds that Google and Broadcom entered a similar arrangement for TPU chips (Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026)). The paper's characterization of a committed guarantee differs from The Information's and the Journal's accounts of an option under memorandums of understanding.

Huang set out his own rationale in an X Article posted August 10, 2026, writing that "we have moved from an era in which companies bought chips and built data centers project by project to one in which AI factories can be financed as productive infrastructure" and that "in AI, compute is revenue." Writing on August 11, 2026, Ben Thompson characterised the up-to-25% residual-value backstop as "in a certain sense, a price cut," on the reasoning that it puts Nvidia's profits behind uncertain investments in order to lower borrowers' cost of capital. He placed the arrangement against a tightening debt market: Oracle, Meta, Alphabet and Amazon raised a combined $108 billion across all of 2025 and, as of July 7, 2026, had already raised $194 billion in 2026, with 86% of the year's bonds trading at higher yields than at issuance and cover on recent issuance down to less than 2x from 5x in February (Source: stratechery.com). See Private Credit & AI Infrastructure, AI Bubble Debate.

Reporting published August 27, 2026 said Nvidia had paused some deals in the financing initiative that offered credit support to AI cloud providers in exchange for a share of their revenue, less than two months after announcing the program. People familiar with the matter said some Nvidia employees had told current and potential customers that the program could draw antitrust scrutiny, and cited sensitivities around the extent to which the chipmaker can dictate how its customers do business. Which deals were paused, and whether the pause is temporary, were not reported (Source: wsj.com; article body behind a paywall, lede only).

A SemiAnalysis analysis published July 6, 2026 detailed the program as a typically six-year take-or-pay minimum-revenue guarantee under which Nvidia shares in Neocloud revenue above the backstop level. It projected outstanding AI debt to exceed $7 trillion by 2029 — second only to the roughly $13 trillion US mortgage-backed market — listed announced backstop deals including SharonAI's 40,000-GB300 Australian project ($4.88 billion, June 2026) and Firmus's 360MW Batam, Indonesia cluster (June 29, 2026), and noted that AMD has offered similar backstops since 2025 (Source: newsletter.semianalysis.com). See Private Credit & AI Infrastructure.

Products and product expansion

Beyond data-center accelerators and discrete GPUs, Nvidia has expanded across the AI stack into CPUs, client silicon, open models, enterprise software, and physical AI.

On its May 20, 2026 earnings call, CEO Jensen Huang described the new Vera CPU as "the world's first CPU, purpose-built for agentic AI," opening "a brand new $200 billion TAM" — Nvidia's first explicit total-addressable-market claim outside the GPU and accelerator product line, repositioning the company as a full-stack vendor for agentic workloads where CPU and GPU pairing matters more than for pretraining-heavy workloads. (Source: finance.yahoo.com) Huang said Nvidia had already sold $20 billion of standalone Vera CPUs during the year and was ramping its Vera Rubin platform. On July 21, 2026, CoreWeave published engineering-sample benchmarks showing the Vera Rubin NVL72 delivering 5.4× performance per megawatt versus the GB200 NVL72 on DeepSeek R1 inference, a figure SemiAnalysis's July 22 analysis put at roughly 2× against the current GB300 baseline (Source: newsletter.semianalysis.com). On May 23, 2026 he clarified that the $200 billion TAM forecast includes China, making the figure partly contingent on either a relaxation of U.S. export controls or Vera-class CPUs falling outside restricted-compute thresholds, given Nvidia's near-total loss of the Chinese accelerator market. (Source: cnbc.com) See AI Coding Agents, Circular Financing in AI, Export Controls (AI).

Nvidia entered the client PC chip market for the first time on May 31, 2026, when Huang unveiled the N1X, an Arm-based PC chip, and — jointly with Microsoft — the first Windows PCs purpose-built for personal AI agents, marketed around an "RTX Spark" platform delivering roughly 1 petaflop of AI performance. Nvidia said it would work with Dell, Lenovo, HP, and Microsoft to build the agentic, on-device AI laptops, putting it into direct competition with incumbent x86 PC-CPU vendors and extending the agentic-AI repositioning from the data center to the laptop. (Sources: cnbc.com; nvidianews.nvidia.com) The launch had been previewed ahead of Computex, when Nvidia and Microsoft teased a consumer-PC announcement under the shared tagline "A new era of PC." Analyst M.G. Siegler argued on May 29, 2026 that the tease pointed to Nvidia's long-rumored general-purpose "N1"/"N1X" CPUs and possibly a paired Microsoft Surface device, characterizing it as a second attempt at the Arm-based Windows PC that failed with the original Surface RT; Siegler labeled his reading analyst speculation rather than a confirmed product. (Source: spyglass.org) On June 1, 2026 the WSJ confirmed the launch as the first PCs designed for AI agents, with Dell, Lenovo, and HP manufacturing the laptops. (Source: wsj.com) See Microsoft, Edge AI / Private Physical AI, AI Coding Agents.

Nvidia released Nemotron 3 Ultra, a 550-billion-parameter open model, around June 2, 2026, in the same early-June window as Google's Gemma 4 12B and Microsoft's MAI slate, extending Nvidia's open-model line as a complement to its accelerator and software stack. (Source: nlp.elvissaravia.com) Nvidia's own research site dates the Ultra release to June 4, 2026 and gives the configuration as 550 billion total with 55 billion active parameters (Source: research.nvidia.com). Nvidia extended the family on August 11, 2026 with Nemotron 3.5 Lightning, a 30-billion-parameter model distilled from the larger Nemotron models and, per Nvidia, able to run on a single GPU in a laptop or desktop (Source: cnbc.com). See NVIDIA Nemotron and Open-Source AI / Open-Weight Models.

Nvidia's investment in a new in-house open-source model family it calls Nemotron 4 became public on August 11, 2026 through an item in The Information, "Why Nvidia Is Trying To Develop The World's Best Open-Source AI Models"; no article body could be retrieved and the publisher's canonical URL was not captured, so the model's scale, timing and licensing are unverified (Source: New Developments Log/2026-08-11-0812-ai-developments.md).

On September 15, 2026, Nvidia and Salesforce jointly announced Koa, Salesforce's CRM reasoning model for its Agentforce platform, built by post-training Nemotron 3 Super with Nvidia's NeMo RL, NeMo Gym and NeMo AutoModel tooling. Salesforce controls Koa's weights and runs it on its own infrastructure; at announcement it was with select pilot customers, with general availability expected in winter 2026. Jensen Huang said in the release that "NVIDIA Nemotron open models give Salesforce the foundation to turn decades of enterprise expertise into specialized AI with Koa." The same announcement brought Nemotron-based models and Nvidia accelerated computing into Salesforce's Missionforce line for government and regulated customers, including air-gapped networks (Source: salesforce.com). Salesforce's technical report states that the model "was trained in partnership with NVIDIA" on a cluster of five Nvidia B200 nodes (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)).

On June 3, 2026, Nvidia disclosed the acquisition of Kumo AI, a five-year-old startup selling predictive AI software to enterprises, for more than $400 million, extending Nvidia's push beyond accelerator hardware into the enterprise model and software layer alongside its client-silicon and physical-AI moves. (Source: theinformation.com)

On July 7, 2026, Nvidia and AI-chip startup d-Matrix disclosed a partnership combining their respective hardware in a new system to power AI models — reported as the latest example of Nvidia's strategy of partnering with, rather than fighting, its growing list of AI server-chip competitors; d-Matrix's Corsair inference accelerator entered volume production on June 9, 2026 (Source: theinformation.com).

Nvidia said at the Hot Chips conference on August 24, 2026 that Groq 3 LPX, a dedicated inference accelerator, had entered full production, with Nebius Group as the first committed customer. The chip is a purpose-built extension to the Vera Rubin platform that offloads token generation from the GPUs handling context ingestion, and a full rack-scale deployment can hold up to 256 LP30 accelerators. In tests run by Artificial Analysis the chip output 3,400 tokens per second on the open-source Gemma 4 31B model with a 100,000-token context window; Nvidia says that makes it four times more responsive than rival platforms for latency-sensitive work. The design rests on technology Nvidia licensed from Groq Inc. for a reported $20 billion in December 2025, a deal that also brought founder Jonathan Ross and president Sunny Madra to Nvidia. Nvidia named SpaceX as a customer for its Vera CPUs, spanning terrestrial data centers and orbital satellites (Source: siliconangle.com). See Specialty Inference Hardware — Cerebras, Groq, SambaNova, Inference Economics and Token Pricing.

On July 15, 2026, Nvidia introduced the Jetson T3000 and T2000 Blackwell modules — rated at 865 and 400 FP4 teraflops respectively — for mass-market robotics and edge AI, alongside Cosmos 3 Edge, a 4-billion-parameter on-device world foundation model; the modules ship in Q1 2027 (Source: blogs.nvidia.com). See Edge AI / Private Physical AI.

Supply and demand for accelerators

Demand for Nvidia accelerators has continued to exceed supply at the frontier. Azeem Azhar's Exponential View documented on May 4, 2026 that Nvidia B200 GPU rental prices grew 114% in six weeks, with the per-hour premium over H200s expanding more than 6x. Lightning AI disclosed that roughly forty of its customers were seeking 400,000 GPUs against its current fleet of about 40,000, and Microsoft began requiring Blackwell customers to lock in at least 1,000 chips for a year. (Source: exponentialview.co) See AI Bubble Debate, Inference Economics and Token Pricing.

Cloud providers tightened access during this period. As of April 24, 2026, Microsoft and other cloud providers were restricting access to Nvidia GPUs, diverting supply to internal teams and large customers such as OpenAI and leaving smaller AI startups struggling. (Source: theinformation.com)

A May 4, 2026 analysis reported that Asian suppliers accounted for 90% of Nvidia's production costs, up from 65% the prior year, reflecting Nvidia's expansion into physical AI (humanoids, autonomous vehicles, edge robotics) where the supplier base for components beyond the Taiwan-fabricated GPU is concentrated in Asia. (Source: fortune.com) See Semiconductor Supply Chain.

SemiAnalysis reported on July 6, 2026 that Nvidia's Kyber rack-scale architecture — a cabinet housing 144 Rubin Ultra chips — has been delayed more than 12 months to 2028 because the PCB midplane "remains challenging from a manufacturability standpoint," and that a fallback dual-rack design was cancelled after pushback from cloud providers, a slip it said could open the high end of the market to AMD and Google; Nvidia rejected the report, saying "Our roadmap is intact" (Source: cnbc.com).

Nvidia notified some of its largest customers of AI server price increases in a move that became public on August 22, 2026, days before its second-quarter earnings. The customers notified build AI data centers for Oracle, Microsoft and others. Accounts of the size differ: Bloomberg and The Verge put the increase above 15 percent, while The Information reported about 17 percent, citing two people whose firms were notified by server makers. The Information's account applies the increase to Grace Blackwell 300 and Vera Rubin 200 chip systems due for delivery in 2027, and estimates that it could raise the cost of a 1-gigawatt data center by at least $5 billion at current system prices (Source: theverge.com; bloomberg.com; Phoebe Liu and Amir Efrati, "Nvidia AI Chip Prices to Rise About 17%, Server Makers Tell Customers," The Information, August 22, 2026 — the newsletter link is a tracking redirect and no canonical URL was captured). See Inference Economics and Token Pricing, AI Data Centers.

During CEO Jensen Huang's June 2026 visit to Seoul, Nvidia announced a cluster of South Korean partnerships on June 8, 2026, including a multi-year memory-supply agreement with SK Hynix that Huang said runs more than two years with options to extend, plus deals under which SK Telecom, Naver, and Doosan would use Nvidia technology to build AI data centers; SK Telecom said it would build a gigawatt-scale AI cloud with its first data center online in 2027. The announcements landed as the Kospi fell almost 9% amid a global tech-stock selloff, which Huang dismissed. (Source: reuters.com)

CUDA software moat

CUDA, Nvidia's proprietary software stack, is widely cited as the company's most durable competitive advantage — more entrenched than its hardware lead — because it locks the AI-developer ecosystem onto Nvidia GPUs. On June 1, 2026, OpenAI executive Sachin Katti signaled openness to open-sourcing internal multi-vendor chip software that could weaken the CUDA moat. A credible, open multi-vendor abstraction layer backed by a frontier lab would lower switching costs to non-Nvidia accelerators (AMD, custom silicon, Huawei). No software had been released as of the disclosure, making this an emerging rather than realized threat. (Source: theinformation.com) See OpenAI, Export Controls (AI).

Nvidia has also released open-source training software of its own. On August 1, 2026 its NeMo team published Molt, a PyTorch-native agentic reinforcement-learning framework, under Apache 2.0, with a technical report at arXiv:2607.21653. Molt comprises roughly 8,600 lines of reinforcement-learning code by import-graph tracing, against about 62,000 for verl and 25,000 for slime, and composes Ray, vLLM and NVIDIA AutoModel with FSDP2 without forking any of them. On Qwen3-30B-A3B across two nodes of eight H100 GPUs, Molt recorded 119.4 ± 2.3 seconds per optimizer step against slime's 109.5 ± 10.3; the report claims no superiority in either direction because the spreads overlap, and discloses an upstream distributed-mixture-of-experts forward mismatch in the benchmark checkpoint (Source: marktechpost.com). See RLHF (Reinforcement Learning from Human Feedback), Agentic AI.

Qualcomm's data-centre push produced a named customer in September. On September 8, 2026 it announced an infrastructure agreement with AWS covering multiple generations of customized inference silicon, disclosing in an SEC filing that it had issued Amazon warrants for 25 million Qualcomm shares at $161.26 each — $4 billion in total — vesting in tranches tied to commercial milestones and to Amazon's purchase of up to $60 billion of Qualcomm server chips and other technology, with the warrant expiring September 3, 2036 (Source: cnbc.com).

A parallel move came from a hardware rival: on June 24, 2026 Qualcomm agreed to acquire AI software startup Modular — developer of the Mojo language and MAX inference platform, both positioned as portable alternatives to CUDA — for nearly $4 billion, framed as strengthening Qualcomm's software stack as it pushes into the data-center market (Source: cnbc.com). The same day, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference accelerator (see OpenAI), part of a broader buildout of non-Nvidia inference silicon. OpenAI published Jalapeño's first measured results on August 25, 2026, benchmarking it on SemiAnalysis's InferenceX against Nvidia GB200 and GB300 systems and reporting 1.5 to 1.9 times more work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency, with all ratios normalized by rated chip power rather than per chip. The same post states that OpenAI will continue to deploy Nvidia accelerators widely for both training and inference (Jalapeño's first results show industry-leading speed and efficiency in AI inference). The figures are OpenAI's own and have not been independently reproduced.

Government and defense contracting

Nvidia is one of the seven companies signed to the Pentagon's classified-network agreements announced May 1, 2026, alongside SpaceX's xAI, OpenAI, Google, Reflection, Microsoft, and AWS; the contracts cover IL6 and IL7 systems. (Source: theinformation.com) See DOD — Department of Defense (AI Deployer).

Nvidia hired veteran lobbyist Bruce Andrews — Intel's former government-affairs chief and an Obama-era Commerce official — as Chief External Affairs Officer on June 11, 2026, as its China sales faced renewed Washington scrutiny. (Source: reuters.com)

Export controls and China

Export controls on Nvidia GPUs to China are the primary mechanism of U.S. compute-centric AI governance. The AI Action Plan calls for strengthening chip export-control enforcement, including location-verification features on advanced AI compute. Nvidia is a founding partner of Project Glasswing.

China was once at least 20% of Nvidia's data-center revenue. On May 21, 2026, following the record Q1 FY2027 quarter, Huang told investors that Nvidia had "largely conceded" China's AI-chip market to Huawei and that they should "expect nothing" on U.S. approvals to resume advanced-chip sales to China. (Source: cnbc.com) Huang repeated the "largely conceded" framing publicly on May 25, 2026, the same day Huawei announced its Tau Scaling Law and LogicFolding at a Shanghai symposium; the statement followed Trump and Xi launching a government-to-government AI dialogue on May 19, 2026. He was also added to President Trump's delegation for an imminent U.S.-China summit, placing Nvidia's CEO inside the diplomatic channel even as he told investors to expect no near-term policy relief. (Source: cnbc.com) See Huawei — Ascend AI Accelerators, Export Controls (AI), AI Race Dynamics.

On May 23, 2026, Huang said it "would be terrific" to serve China with the H200 chip and disclosed that the U.S. had licensed H200 sales to roughly 10 Chinese firms, though no deliveries had been made and Chinese regulators had not approved them — the first numerical disclosure of how many Chinese firms hold U.S. H200 export licenses. (Source: reuters.com) The stalemate showed signs of easing in July: per reporting published July 8, 2026, China planned to allow some of its biggest AI companies to buy limited quantities of H200 chips to offset a domestic shortage caused by soaring demand — the roughly ten U.S.-cleared firms had still taken no deliveries amid Beijing's guidance against foreign chips (Source: theinformation.com). Amid the July 2026 controversy over alleged Chinese distillation of U.S. models, Huang defended Chinese open models in a July 21, 2026 interview — "There's no scenario where China runs U.S. companies off the road" — and urged Anthropic to make Claude Mythos broadly available (Source: axios.com). See Distillation.

With its China GPU share fallen to roughly zero amid stalled H200 deliveries, Nvidia pivoted to selling its new Vera CPU to Chinese clients: on June 12, 2026 it told them the CPU could ship as soon as August and that they could order immediately, with one cloud company planning an order of more than 300 dual-CPU servers. A single Vera was priced "well north" of $20,000, and Nvidia expected $20 billion in Vera revenue in the fiscal year. (Source: reuters.com) The China-Vera push tests whether CPU-class silicon falls outside the restricted-compute thresholds that have shut Nvidia's accelerators out of the market.

Diversion enforcement tightened on the seller side in July 2026: per Financial Times reporting of July 13, Nvidia more than halved its list of Asian customers authorized to buy advanced AI chips, introducing a "white list" with tougher vetting of buyers in Singapore, Malaysia, and Japan to close loopholes used to route chips to China (Source: ft.com). In the opposite direction, the U.S. loosened export controls on the UAE on July 10, 2026, easing Nvidia AI-chip sales; the Gulf state's flagship AI firm G42 can now freely buy chips — described in reporting as a reward for UAE support of the U.S. war in Iran — and plans to become a U.S. company (Source: wsj.com; investing.com). See Export Controls (AI).

Reporting published August 20, 2026 described a route into China separate from the H200 licences. According to two Nvidia employees, the company plans to begin small-batch shipments of an AI chip specially tailored for Chinese customers by the end of 2026. The chip is a variant of Nvidia's language processing unit — a chip developed with technologies licensed from Groq that works alongside Nvidia GPUs to speed up a chatbot's response times — and several Chinese customers had already placed orders. Only the article's opening two paragraphs were retrievable, so the chip's specifications, its position relative to US export thresholds, and whether it requires a licence are not established (Source: theinformation.com).

Speaking in Taipei on May 23, 2026, Huang urged longtime partner Super Micro to tighten compliance after Taiwan detained three people in its first crackdown on semiconductor smuggling, tied to allegedly fraudulent declarations about AI servers bound for China, exposing U.S. chip vendors' OEM partners to Taiwanese liability. (Source: bloomberg.com) See Export Controls (AI).

That investigation reached Nvidia's own staff. Taiwan's Keelung District Prosecutors Office said on July 28, 2026 that it had detained a suspect surnamed Chang — identified by Mirror Media and Bloomberg as an Nvidia employee — on suspicion of forgery and breach of trust over AI servers exported to China in violation of U.S. export controls. Investigators searched his home and workplace, including a desk at Nvidia's Taipei office, on July 24, 2026. Seven people are in custody in an investigation opened in May 2026 that centres on roughly 50 falsely documented Super Micro servers. The prosecutors' statement does not accuse Nvidia of wrongdoing, and Nvidia said smuggling is a "nonstarter" (Source: qz.com; forbes.com). See Chip Smuggling and Export-Control Evasion.

Taiwanese prosecutors' indictment of nine people, including employees of Nvidia and Super Micro, over the alleged illegal export of 74 high-end AI servers to China became public on August 24, 2026. No article body was retrievable and the charges are not verified beyond the headline; the count of nine and the figure of 74 servers both exceed the seven detentions and roughly 50 servers recorded in the July stage of the same investigation (Source: reuters.com).

Huawei rotating chairman Eric Xu said on September 17, 2026 that reliable data on Nvidia's China share was difficult to obtain but that "I think Ascend market share should be bigger than Nvidia's"; Huawei provided no supporting data (Source: reuters.com). See Huawei — Ascend AI Accelerators.

The RTX Pro 5500 became the next candidate for Chinese sales. According to two people familiar with the matter, in an account that became public on September 27, 2026, China's Ministry of Industry and Information Technology asked companies including Alibaba Group and ByteDance, ahead of Xi Jinping's Washington visit, to report how many of the chips they want and for what use, and told some companies the government intends to approve the purchases; the timing, quantity and standard for approvals were unclear, and the U.S. government had not publicly addressed whether the chip falls outside export restrictions. ByteDance is weighing an order of about 1 million. Nvidia aims to start shipping in late December and targets about 500,000 chips a quarter for China, and its sales staff told customers to order by September 30. Two chip buyers put the chip's price in China at 85,000 to 90,000 yuan (about $13,000), roughly that of Huawei's Ascend 950PR; the China-specific RTX Pro 6000D sells for about 65,000 yuan, up from about 50,000 yuan the previous year. H200 sales to Chinese buyers were less than 1% of Nvidia's data center revenue in the quarter ended July 26, 2026 (Source: theinformation.com).

People

On July 2, 2026, Nvidia recruited 26-year Microsoft veteran Nick Parker, who led Microsoft's worldwide commercial sales business, as executive vice president of worldwide field operations, with a pay package exceeding $40 million (Source: geekwire.com).

Chris Malone, OpenAI's former head of data centers, joined Nvidia in September 2026 as vice president of its DSX Platform, the division that helps customers design and build AI data centers to Nvidia's specifications, according to his LinkedIn profile as reported by The Information; the move became public on September 25. Malone had left OpenAI in August. Only the article's summary was retrievable (Source: theinformation.com).

Positions on open-weight AI

Nvidia signed the July 24, 2026 industry letter "Open Weights and American AI Leadership" (Open Weights and American AI Leadership (industry letter, July 2026)), which Jensen Huang marked with his first post on X, warning the industry against repeating a mistake he says software narrowly avoided in the 1980s (Source: fortune.com).

According to people familiar with the matter in a Wall Street Journal account of August 22, 2026, Nvidia plans to use the $6 billion non-exclusive licensing deal it struck with Poolside to build one of the world's most powerful open-weight AI models, positioned against Chinese open-weight models including DeepSeek's and Kimi K3, and against OpenAI and Anthropic on the grounds that open-weight models are cheaper to operate and easier to customize. The article's remainder sits behind a subscription wall and no further terms could be read (Source: wsj.com).

On July 27, 2026 the company launched the Open Secure AI Alliance with more than 25 inaugural partners, calling on regulators to treat open models as "defensive assets, not liabilities" and contributing the NVIDIA Labs Object-Oriented Agent framework as open source. Nvidia's launch argument cited the July 2026 Hugging Face intrusion: closed commercial models refused the forensic work, so Hugging Face ran open-weight GLM 5.2 on its own infrastructure to analyze more than 17,000 recorded actions (Source: blogs.nvidia.com; huggingface.co). See Open-Weight Frontier Models.

Positions on AI and the labor market

On the May 1, 2026 Dwarkesh Patel podcast, Huang called rival CEOs' job-loss predictions "hurtful" and the work of leaders with "a God complex," pushing back specifically on Dario Amodei's forecast of "50% of entry-level white-collar jobs in five years," Mustafa Suleyman's "18 months," and Elon Musk's "end of all human jobs." Demis Hassabis also publicly rejected the "bloodbath" narrative. In the same interview, Huang said Anthropic's Mythos was "trained on fairly mundane capacity, and a fairly mundane amount of it," a comment Matt Stoller's May 4 column read as undermining the U.S. compute-moat thesis underpinning hyperscaler capex. (Sources: fortune.com; thebignewsletter.com) See AI Labor Disruption, Claude Mythos Preview.

Investments and licensing deals in AI companies

Nvidia opened discussions to invest in Perplexity as part of an equity round valuing the company at more than $30 billion, reported on August 23, 2026. The round would be worth billions of dollars. Nvidia had previously considered paying Perplexity billions of dollars to license its technology and hire staff, the structure it used with Groq, Enfabrica and Poolside; what prompted the shift toward a conventional equity investment is not established. Perplexity joined Nvidia's Nemotron Coalition in March 2026, and representatives of the two companies are reported to meet several times a week on hardware and software development (Source: theinformation.com).

The Groq structure drew a federal antitrust inquiry. A Justice Department investigation into whether Nvidia structured the licensing deal to avoid antitrust scrutiny became public on September 9, 2026; the department opened it shortly after the December announcement and has sent Nvidia a formal request for information. An Nvidia spokesperson said "the Groq story is a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers." The department could fine Nvidia if it finds the deal was mishandled but is unlikely to seek to unwind it (Source: reuters.com). That account values the deal at $17 billion, against the roughly $20 billion reported at announcement; the divergence is unresolved. See Groq, AI Antitrust.

Equity stakes in power and siting firms (August 2026)

Nvidia took a series of minority equity positions in firms controlling the land, power and orbital capacity that data centers depend on. It said on August 17, 2026 that it would invest $1.5 billion in SoftBank-owned SB Energy. On August 21, 2026 Nvidia and Cloverleaf Infrastructure said Nvidia had made a minority investment in the privately held firm, which works with utilities, energy providers and investors to secure power and land for data-center sites; terms were not disclosed, though the Wall Street Journal reported earlier that day that the figure was expected to run to several hundred million dollars. Cloverleaf will deploy Nvidia's DSX platform to shape site selection, power, cooling and computing decisions, and says it has delivered multiple gigawatt-scale projects since its founding in 2024. Its chief executive Brian Janous, a former Microsoft vice president for energy, said the firm has a pipeline of 10 GW of projects across the Southeast and Midwest and that "right now we do have a chip overhang problem" caused by power delays (Source: reuters.com; theinformation.com).

The same day, orbital-inference startup Starcloud said it had added a $250 million extension to the $170 million Series A it raised in March 2026, valuing it at $2.3 billion. Manhattan West Ventures led the extension with participation from Nvidia and Cisco; a person familiar with the deal said Nvidia contributed $25 million (Source: techcrunch.com). See Data Center Siting / AI Power Politics, AI Data Centers.

Grid flexibility and the AI Energy Management Alliance

Nvidia has promoted data centers that adjust their electricity use to grid conditions as a route to faster grid connections. On March 23, 2026, at CERAWeek, Nvidia and Emerald AI announced that they were working with AES, Constellation, Invenergy, NextEra Energy, Nscale Energy & Power and Vistra on AI data centers built to Nvidia's Vera Rubin DSX reference design, which includes a DSX Flex software library for connecting facilities to power-grid services. The release said the two companies had trialed power-flexibility demonstrations at five commercial data centers over the previous year and that DSX Flex was expected to be deployed at commercial scale later in 2026 at the Nvidia AI Factory Research Center in Virginia. Jensen Huang said that "every system must be designed together — energy, compute, networking and cooling as one architecture" (Source: nvidianews.nvidia.com). Nvidia participated in Emerald AI's $150 million Series A, announced August 25, 2026 at a $1.05 billion valuation (Source: businesswire.com).

On September 16, 2026 Nvidia, Google and Emerald AI launched the AI Energy Management Alliance as founding members, with 18 launch partners including Anthropic, National Grid, AES, Constellation, NRG and RWE. Josh Parker, Nvidia's head of sustainability, said Nvidia joined "to advocate for technology-neutral, performance-based standards that evaluate flexible data centers on measurable grid reliability metrics," with the result of "faster AI deployment and a more efficient grid, without asking ratepayers to fund new capacity" (Source: businesswire.com). Parker described the approach to reporters as "a win-win-win solution" (Source: axios.com).

Data-center leases behind customer cloud contracts (August 2026)

Nvidia holds the lease on the Nueces County, Texas data center behind the cloud-computing agreement between Lambda and Anthropic reported at $35 billion, which became public on 31 August 2026. Nvidia is also to supply the chips. The campus is being developed by Hut 8, a bitcoin miner and data-center developer, and Nvidia signed an agreement with Hut 8 a few weeks earlier to secure the capacity (Source: wsj.com). The arrangement extends the pattern of the equity positions above from land and power into the leases underlying its customers' own cloud contracts.

An investment by Nvidia of $3.5 billion in convertible bonds issued by MediaTek, alongside an arrangement for MediaTek to produce Nvidia's NVLink Fusion chiplets, switches and custom memory, became public on 31 August 2026. The report rests on a single subscription briefing whose body could not be retrieved and is not verified beyond that account (Source: theinformation.com).

TSMC

Overview and role

TSMC fabricates almost every leading AI chip; Nvidia, AMD, Apple, Qualcomm, and others all depend on its advanced process nodes. The United States hosts 5,427 data centers (10x any other country), but the chips inside them are overwhelmingly fabricated by TSMC in Taiwan (Stanford HAI AI Index Report 2026). A TSMC U.S. expansion began operations in 2025, partially reducing but not eliminating the Taiwan dependency.

Capacity, demand, and pricing

At TSMC's annual shareholder meeting in Hsinchu on June 4, 2026, CEO C.C. Wei warned that AI-driven chip shortages could persist for years. Even with six planned U.S. facilities representing $165 billion in investment, "it will be a long time before we can meet customer demand." Wei reiterated guidance of more than 30% sales growth and said TSMC would not pursue abrupt price hikes, against a backdrop of roughly $725 billion in expected hyperscaler AI spending during the year. (Sources: tradingview.com; sdxcentral.com)

On July 12, 2026, Taiwan's National Science and Technology Council Minister Wu Cheng-wen said TSMC will add a third and fourth advanced chip-packaging plant at the Chiayi Science Park — projected at more than NT$300 billion (about $9.35 billion) in annual production value and over 9,000 jobs — as demand for chip-on-wafer-on-substrate (CoWoS) capacity from AI chip designers such as Nvidia outstrips supply (Source: reuters.com).

On July 16, 2026, TSMC said it will invest an additional $100 billion in U.S. manufacturing, adding at least four chip plants and advanced-packaging facilities in Arizona and bringing its announced U.S. total to $265 billion, citing surging AI infrastructure demand (Source: theinformation.com). The commitment expands the six-facility, $165 billion U.S. program CEO C.C. Wei described at the June 2026 shareholder meeting; Bloomberg characterized the buildout as central to TSMC's arrangement with the Trump administration (Source: bloomberg.com). The announcement came alongside second-quarter results: revenue up 36% year over year to roughly $39.45 billion and net income up 77% to a record of about $22 billion, with 2026 capital-expenditure guidance raised to $60–64 billion. Wei cited "the AI megatrend" and said AI demand looks "very strong... all the way to probably 2029, 2030" (Source: cnbc.com; fortune.com).

As TSMC's leading-edge capacity tightened, Google and Nvidia were reported on June 8, 2026 to have held talks about using Intel's contract-manufacturing arm as a backup, with Google having ordered more than 3 million Tensor Processing Units from Intel for 2028 and Nvidia assessing whether Intel's process could support a design combining four graphics chips in one package. (Source: finimize.com) See Intel.

Geopolitical risk

If China invaded Taiwan, the West would lose access to its primary source of frontier AI chips; Toby Ord identifies this as a key risk in long-timeline worlds, noting that by 2035 China may have invaded Taiwan. The Stanford HAI AI Index 2026 frames the dependence on a single foundry as "a single-point-of-failure vulnerability" (Stanford HAI AI Index Report 2026). The AI Action Plan's CHIPS program aims to restore American semiconductor manufacturing to reduce this dependency.

CoWoS and HBM chokepoint

According to SemiAnalysis, Nvidia is not fab-constrained but packaging- and memory-constrained (Source: Raw Sources/SemiAnalysis - CoWoS and HBM Supply Chain.md):

• TSMC CoWoS capacity: approximately 35k wafers/month (2024), rising to about 70k (2025) and a 110k target (2026), oversubscribed through 2026. Every Nvidia H-, B-, and GB-series accelerator requires CoWoS.
• HBM: SK Hynix, Samsung, and Micron have all sold out 2026 HBM output under long-term contracts, with shortages extending into late 2027. SK Hynix is Nvidia's dominant HBM supplier (more than 95% HBM3 share at peak).

Nvidia's ability to ship more B200 and GB200 accelerators is therefore gated by TSMC backend capacity and HBM allocation, not by wafer starts at 4NP. C.C. Wei's June 2026 multi-year shortage warning is the supply-side counterpart to demand-side capex figures, with the binding foundry conceding it cannot close the gap on a multi-year horizon. See Semiconductor Supply Chain for the full chokepoint mapping.

Related concepts

• Semiconductor Supply Chain — the full chokepoint map; CoWoS and HBM are the binding constraints
• Compute Governance — Nvidia and TSMC are the physical substrate that compute governance governs
• Export Controls — chip export restrictions are the primary tool
• AI Sovereignty — compute access is the binding constraint on national AI sovereignty
• Scaling Laws — Nvidia and TSMC control the supply of compute
• Huawei Ascend — the Chinese alternative compute stack

Relationships

• related: The Nvidia-sized hole in US GDP statistics (Epoch AI, August 2026) — Epoch AI report arguing Nvidia's US value-add is largely missing from US GDP
• related: Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026) — BPEA paper citing the $500 billion financing platform and 25% residual value guarantee as vendor-finance circularity
• deployed-by: OpenAI, Anthropic, Google DeepMind, Meta AI, Microsoft, Amazon, CoreWeave, Lambda Labs, Poolside — the developers and cloud operators whose compute this page's supply analysis sits underneath
• regulated-by: the Bureau of Industry and Security and the export-control regime described in Export Controls (AI) — chip export restrictions are the primary policy instrument acting on both companies
• depends-on: Semiconductor Supply Chain — the CoWoS and HBM chokepoint section is a specific case of the general map
• supports: Compute Governance — Nvidia and TSMC are the physical substrate that makes compute a governable input; the concept's feasibility rests on the concentration this page documents
• supports: Scaling Laws — the supply of compute is what makes the scaling relationship actionable
• contradicts: AI Bubble Debate — the demand and revenue figures in the Nvidia section are cited on both sides of that debate; this page records the evidence rather than a verdict
• related: Intel and Hugging Face
• related: NVIDIA Nemotron — Nvidia's own model family; Salesforce Koa — Salesforce model post-trained from Nemotron 3 Super and announced jointly with Nvidia
• related: Jensen Huang
• related: AI Energy Management Alliance (AEMA), Emerald AI — founding member of the coalition; partner in and investor in Emerald AI
• related: AI Sovereignty, AI Data Centers, Inference Economics and Token Pricing, Circular Financing in AI, Private Credit & AI Infrastructure, Edge AI / Private Physical AI, AI Coding Agents
• related: Stanford HAI AI Index Report 2026, America's AI Action Plan, Open Weights and American AI Leadership (industry letter, July 2026), Jalapeño's first results show industry-leading speed and efficiency in AI inference