The single greatest systemic vulnerability is the AI ecosystem’s shared, concentrated fabric: a handful of compute chokepoints, a shared data commons, a shared regulatory environment,...

Strategic analysis · August 2026, by Claude

A collapsing AI incumbent has a large arsenal of “salt-the-earth” instruments. The most credible are dumping model weights to destroy rivals’ pricing power; intellectual-property and copyright litigation that imposes cost and legal uncertainty; predatory inference pricing; and exploiting the deeply interdependent capital structures — Microsoft–OpenAI, Amazon/Google–Anthropic, Nvidia–OpenAI — whose “circular” financing means one firm’s collapse mechanically damages its backers.

The single greatest systemic vulnerability is not any one weapon but the AI ecosystem’s shared, concentrated fabric: a handful of compute chokepoints (Nvidia and CUDA, TSMC advanced packaging, high-bandwidth memory), a shared data commons, a shared regulatory environment, and round-tripped balance sheets. A spiteful actor does not need to build a new weapon — it can pull a thread in this shared fabric.

Most scorched-earth moves are self-limiting: they are slow (litigation takes years), legally constrained (predatory-pricing and antitrust law), or self-harming (you cannot credibly “poison the regulatory well” without also being the target). The genuinely dangerous residuum is the class of one-shot, irreversible defections — open-sourcing crown-jewel weights, provoking a safety incident that triggers an industry-wide crackdown, and triggering cross-default or convertible-note cascades in the circular financing web.

The scenario is one of spite, and it is worth distinguishing from adjacent ideas, because the distinction determines which instruments are credible.

Deterrence and mutually-assured destruction (MAD). Cold-War MAD is a stabilising equilibrium: mutual second-strike capability means neither side attacks, and the winning move is not to play. The AI analogue is mutual patent portfolios or cross-licences that deter litigation — detente, not conflict.

Scorched earth. A retreating army destroys everything of value so the advancing enemy cannot use it. The defining feature is that the destroyer has already conceded the ground. Applied here: a lab that knows it will lose destroys the profitability — or the safety — of the market for everyone.

Spite and “salting the earth.” The purest form of the scenario: permanent, irreversible denial of value with no benefit to the actor.

Why credible collapse changes incentives. A healthy firm cannot credibly threaten scorched earth, because the threat would harm itself — a commitment problem. A collapsing firm can, precisely because its “little to lose” status makes the threat credible. This is the core danger, and it is why the instruments below are ranked partly by how irreversible and one-shot they are: irreversibility is what makes a spite move both credible and dangerous.

The twelve categories below enumerate the strategic levers available, with the real-world precedents that show each is more than theoretical. They run roughly from the most conventional (litigation) to the most systemic (defecting from shared norms and triggering financial contagion).

Copyright and training-data litigation. The live example is The New York Times v. OpenAI and Microsoft (filed December 2023, Southern District of New York). On 4 April 2025 Judge Sidney H. Stein denied OpenAI’s motions to dismiss the core direct-infringement claims and the contributory-infringement and trademark-dilution claims, sending the case toward discovery that will force disclosure of training-data practices; the matter is now consolidated as a multidistrict litigation. Separately, the Anthropic copyright class action, Bartz v. Anthropic, settled for 1.5 billion dollars — roughly 3,000 dollars per work across a list of 482,460 books — with final approval granted on 20 July 2026 by Judge Araceli Martínez-Olguín, described by plaintiffs’ counsel as the largest copyright recovery in history. An earlier ruling held that training on books could be fair use, but that downloading pirated copies for a permanent library was not. A collapsing content-rich incumbent, or a data-owning ally, could weaponise copyright by refusing to settle, seeking destruction of training datasets, and driving maximum discovery cost.

APIs and copyright as a structural weapon. Oracle v. Google, over 37 Java API packages, ran from 2010 to 2021, with Oracle seeking as much as 9 billion dollars. The Supreme Court ultimately sided with Google on fair use, but the case imposed a decade of legal uncertainty on Android. The lesson: even a losing suit can impose years of cost and strategic uncertainty — litigation as the weapon, independent of outcome.

Patent thickets and consortia. The smartphone wars featured patent-buying consortia (the Rockstar group’s acquisition of Nortel’s portfolio to sue Android makers), Apple v. Samsung, and Nokia’s campaigns. A dying AI incumbent could sell or weaponise its patent portfolio to a patent-assertion entity — submarine patents, standard-essential-patent holdup — to burden rivals after the operating business is dead.

Trade-secret, tortious-interference, and contract suits. The mutual litigation between Elon Musk and OpenAI shows the template, with OpenAI’s countersuit explicitly alleging that Musk’s attacks damaged its ability to recruit talent and close deals. Litigation here is framed openly as competitive interference.

Feasibility and counters. Litigation is slow and expensive, and courts have so far leaned toward fair use on training itself. Its value as a scorched-earth tool is cost-imposition and uncertainty, not victory: highly credible because cheap relative to the balance sheets involved, but slow and self-limiting.

Reverse acqui-hires are the signature AI-era instrument. Microsoft’s 2024 Inflection deal (around 650 million dollars, structured as a non-exclusive model licence plus a payment to waive claims over the hiring of staff) gutted the startup while leaving a hollow shell — structured to avoid premerger antitrust notification. The pattern repeated across Amazon–Adept, Amazon–Covariant, Google–Character.AI, Google–Windsurf, and Meta–Scale AI, drawing antitrust inquiries on both sides of the Atlantic.

The scorched-earth variant. A collapsing lab dumping its entire research team onto the market — or into a single competitor — either floods the labour market, depressing the value of a rival’s own retention packages, or hands one favoured rival a decisive talent windfall while denying others. Non-compete and non-solicit litigation and poaching of key researchers are the friendlier-fire versions.

Feasibility and counters. Highly credible and already routine. Antitrust scrutiny of quasi-mergers is rising but uneven, and California bars most non-competes — so talent liquidity is a genuine systemic vulnerability.

This is arguably the most credible and most dangerous scorched-earth instrument, because it is fast, one-shot, and irreversible: once weights are public, they cannot be recalled.

The live strategy. Meta’s Llama releases are a textbook “commoditise your complement” play — by driving the price of the model layer toward zero, Meta erodes the pricing power and API margins of OpenAI, Anthropic, and Google. Critics describe it as: if you cannot win the proprietary game, commoditise the model layer so nobody else can win it either.

The demonstration of destructive power. DeepSeek’s open-weight R1 (released January 2025, reportedly trained for under 6 million dollars) triggered the largest one-day market-cap loss in history. Per Reuters, citing LSEG data, Nvidia lost 593 billion dollars on 27 January 2025 — shares down roughly 17 percent, their worst day since March 2020 — alongside Broadcom down 17.4 percent, Alphabet down 4.2 percent, and the Philadelphia semiconductor index down 9.2 percent. A single open release repriced the entire AI value chain.

The scorched-earth endgame. A dying lab could open-source its crown-jewel frontier weights and training recipes purely to destroy the market’s profitability — converting its most valuable asset into a free public good specifically so that no surviving rival can earn monopoly rents on comparable capability. This is the equivalent of a retreating army giving away the armoury.

Feasibility and counters. Extremely feasible and irreversible. Partial counters: releasing weights may violate data or licensing agreements or trigger regulatory review of dangerous-capability release, and the released model still needs compute to run, so it does not destroy the compute layer’s value.

The live price war. By 2026 the industry is in open price competition, with inference prices falling roughly an order of magnitude per year, aggressive free tiers, and low-cost open providers setting a price floor that incumbents must match or cede the cost-sensitive market.

Precedents. Platform and subsidy wars — ride-hailing, cloud storage, streaming — show the pattern: the player with the deepest balance sheet subsidises below-cost service to bleed competitors who cannot sustain losses.

The scorched-earth variant. A well-capitalised incumbent, or one willing to deficit-spend on its way down, gives inference away free or below cost specifically to ensure rivals that lack a hyperscaler balance sheet cannot survive the margin compression. The goal is not to win customers but to make the market unprofitable for everyone.

Feasibility and counters. Credible for balance-sheet-rich players who can cross-subsidise from advertising or cloud profits. But predatory-pricing law — which requires below-cost pricing plus a probability of recoupment — is a real constraint, and cheaper inference can expand the market rather than destroy it, blunting the weapon.

The chokepoints. A handful of physical bottlenecks concentrate the ecosystem: Nvidia’s priority allocation of TSMC advanced-packaging capacity (the binding constraint on GPU output) and of high-bandwidth memory; the CUDA software moat supporting data-center gross margins in the mid-70-percent range; and, increasingly, data-center power and grid interconnection. Allocation itself is a weapon: when compute is the binding constraint, whoever controls allocation controls who can ship.

Scorched-earth variants. A well-capitalised incumbent locks up long-term GPU contracts, foundry capacity reservations, memory supply, and data-center power to starve rivals of compute; exclusive cloud-capacity deals deny rivals capacity; or a collapsing infrastructure player breaks or dumps long-term supply contracts, throwing allocation into chaos. Energy is the new bottleneck — multi-gigawatt projects show that compute is now gated by power and interconnection queues, which are themselves fragile.

Feasibility and counters. Locking up supply is highly credible for the cash-rich but expensive. Hyperscaler custom silicon (Google’s TPU, Amazon’s Trainium, Microsoft’s Maia) is actively eroding the Nvidia and CUDA monopoly, and packaging capacity is expanding — so chokeholds are eroding, not strengthening, over a three-to-five-year horizon.

Exclusive data licensing as denial. Google’s roughly 60-million-dollar-a-year Reddit deal is the template, and publishers are erecting bot paywalls. A data-rich incumbent, or a social platform allied to one, could sign exclusive licences to wall off high-value training corpora — Reddit, Stack Overflow, X, news archives — from rivals, or a collapsing platform could sell exclusive rights purely to deny them to competitors.

Data poisoning (systemic risk, not a recipe). Academic research shows that poisoning a tiny fraction of a web-scale corpus is technically feasible. This is treated here strictly as a systemic vulnerability of the shared data commons: the point is that the commons all labs draw from is a shared, corruptible resource, and its degradation — including by AI-generated slop contaminating future training data — harms everyone.

Feasibility and counters. Exclusive licensing is highly credible and legal, and is the most straightforward denial instrument. Deliberate poisoning of shared corpora is legally radioactive, detectable at scale with provenance tracking, and would poison the actor’s own models — so it is largely theoretical as a deliberate corporate weapon, though a real accidental and systemic risk.

The regulatory-moat critique. The core argument is that incumbents advocate safety regulation — licensing regimes, compute thresholds, pre-deployment testing and approval — calibrated so they can comply but startups and open-source projects cannot, converting existing compliance overhead into a legal barrier to entry. The 2024 fight over California’s SB 1047 crystallised this: supporters included Anthropic (after amendments) and leading researchers; opponents included OpenAI, Google, Meta, and prominent venture firms, who warned it would burden all developers and chill open source. The EU AI Act’s obligations for general-purpose models are the live global example.

Weaponising antitrust and investigations against rivals. Filing antitrust complaints, encouraging regulatory probes, and export-control lobbying are all instruments to impose cost and constraint on rivals.

Feasibility and counters. Highly credible and cheap relative to balance sheets. But the regulatory environment is volatile and can cut against would-be capturers — a policy pivot toward voluntary, rather than mandatory, testing and licensing shows the moat can be filled in by the same government that dug it.

Safety-washing and reputational branding. Using “AI is dangerous” framing to brand rivals as reckless, trigger moratoria, or justify a regulatory moat. The critique is that even sincere safety advocacy conveniently proposes the safety regime incumbents clear most easily.

“Burning the commons” (incentive-level analysis only). The genuinely scorched-earth variant is a defecting actor deliberately provoking a dramatic public incident or reckless capability release to trigger an industry-wide crackdown — a moratorium, licensing freeze, or liability shock — that harms all players, including the survivors. Recent episodes in which a single safety incident led to fast, industry-wide access restrictions show how live this channel is. A collapsing actor with little to lose could rationally prefer to trigger such a shock. This document describes the incentive structure and precedent only; it provides no operational pathway.

Feasibility and counters. The narrative weapons are highly credible and already deployed. The incident variant is the single most dangerous theoretical move because it is one-shot and can harm all rivals at once — but it is heavily self-limiting: the provocateur is the most obvious target of the resulting crackdown and liability, so it is credible only for an actor that has genuinely given up.

Instruments. Coordinated public relations, funding critical research and think tanks, whistleblower campaigns, strategic leaks, astroturfing, and sowing distrust in rivals’ model safety or data practices. OpenAI’s tortious-interference countersuit against Musk — alleging a campaign of obstruction designed to damage recruiting, fundraising, and operations — is a live example of reputational attack being litigated as competitive harm.

Feasibility and counters. Cheap, fast, and highly credible; the main constraint is defamation and tortious-interference liability, which cuts both ways, and the difficulty of controlling narrative blowback.

Bundling and distribution moats. Microsoft embedding OpenAI models across Office, Windows, and developer tooling; Google integrating Gemini across Android, Chrome, and Search; Apple Intelligence; Amazon steering cloud customers to Anthropic and Trainium. Distribution is a moat rivals cannot easily replicate.

API, format lock-in, and interoperability denial. Proprietary APIs, model formats, and agent protocols create switching costs; denying interoperability — or breaking it on the way down — strands rivals’ integrations. CUDA is the canonical lock-in precedent.

Feasibility and counters. Highly credible for the platform owners. The counter is the open-weights movement and multi-model routing — enterprises deliberately staying provider-agnostic — which erodes lock-in.

This is the category the scenario makes most consequential, because the AI ecosystem’s capital structure is uniquely interdependent — investment is round-tripped back to the backer as compute spend.

The interdependent structures. After a 2025 restructuring into a public benefit corporation, Microsoft holds roughly 27 percent of OpenAI (valued around 135 billion dollars on about 13.8 billion invested), and OpenAI committed to purchase an incremental 250 billion dollars of Azure services. Amazon has invested around 8 billion dollars in Anthropic with an agreement for up to 25 billion more, and Anthropic committed to spend over 100 billion dollars on AWS and Trainium. Google holds an equity stake (court-confirmed and capped, with no board seat) plus a commitment of up to 40 billion more and a reported 200-billion-dollar, five-year cloud and TPU commitment. Nvidia announced an “up to 100 billion dollar” pledge to OpenAI — explicitly tied to OpenAI buying or leasing Nvidia chips, with OpenAI’s finance chief noting most of the money would go back to Nvidia — though the plan was later reported to have been scaled back to a roughly 30-billion-dollar equity investment.

Why this is a detonator. These deals are circular: the backer’s investment returns as the recipient’s compute spend, and the recipient’s soaring valuation inflates the backer’s reported earnings. Amazon’s first-quarter 2026 filing states verbatim that net income included pre-tax gains of 16.8 billion dollars from its Anthropic investments — more than 40 percent of pre-tax income — even as free cash flow collapsed to 1.2 billion from 25.9 billion a year earlier. Per The Information, contracts involving Anthropic and OpenAI now account for more than half of the roughly 2 trillion dollars in cloud-revenue backlog across Amazon, Microsoft, Google, and Oracle, with Anthropic’s 200-billion-dollar Google commitment alone representing more than 40 percent of Google’s disclosed backlog. The collapse of one large model lab would therefore mechanically blow a hole in a backer’s earnings, evaporate a huge chunk of a hyperscaler’s cloud backlog, trigger convertible-note conversions or defaults, and potentially cascade to the compute suppliers. No single actor designed this contagion channel, but a collapsing actor could exploit it — by triggering cross-default clauses, dumping stakes, calling convertible notes, or litigating partnership terms on the way down.

Feasibility and counters. Highly credible and uniquely dangerous precisely because the structures already exist. Counters: the deals include antitrust caps (sub-controlling stakes, no board control), independent-verification clauses, and diversification (Anthropic runs on TPU, Trainium, and Nvidia; OpenAI moved off Azure exclusivity). But the sheer scale of round-tripping is a genuine systemic fragility.

Defecting from voluntary commitments. Industry coordination — the Frontier Model Forum, voluntary government commitments, shared safety funds, and codes of practice — is largely self-enforced. A defecting actor that breaks these pacts, racing to release, refusing third-party testing, or abandoning safety evaluations, forces rivals into a race to the bottom: once one actor cuts corners for speed, all others face pressure to match or lose the race. This is the classic racing-dynamics collective-action failure.

Feasibility and counters. Defection is trivially easy because the commitments are voluntary and unevenly honoured. The counter is that governments are converting voluntary norms into harder law, which raises the cost of defection.

Ranked by how credible and how irreversible they are, the instruments sort into three tiers.

Most dangerous — one-shot and irreversible. Open-sourcing crown-jewel weights; provoking a safety incident that triggers an industry-wide crackdown; and triggering financing cascades (cross-defaults, note calls, stake dumps) in the circular capital structure. These are the moves a firm with nothing left to lose can execute quickly and cannot walk back.

Credible but self-limiting. Litigation, exclusive data lock-ups, regulatory capture, and predatory pricing. All are real and already deployed as ordinary competition, but each is slow, legally constrained, or self-harming, so their scorched-earth value is cost-imposition and delay rather than a knockout.

Mostly theoretical. Deliberate poisoning of shared corpora and outright sabotage of shared infrastructure. These are legally radioactive and technically detectable, and they poison the actor’s own position — so they matter more as accidental or systemic risks than as deliberate weapons.

The unifying point: the field’s vulnerability is its shared fabric — concentrated compute, a shared data commons, a shared regulatory environment, liquid talent, and interdependent balance sheets. A spiteful actor rarely needs a bespoke weapon; it needs only to pull a thread everyone is standing on.

• Investors and backers: stress-test the circular-financing web. Model the balance-sheet impact of a single large-lab failure on Amazon (the multi-billion-dollar quarterly Anthropic mark-to-market gain reverses), Google, Nvidia, and Microsoft. Treat concentration of more than half of the roughly 2-trillion-dollar cloud backlog in about two labs as a systemic-risk flag.

• Rival labs: diversify compute across at least two chip families and two clouds; avoid exclusive data dependencies; keep a defensive patent and prior-art portfolio.

• Policymakers: continue scrutinising quasi-mergers and reverse-acqui-hires, and require transparency on circular financing and cloud-commitment concentration.

Watch for benchmarks that should change posture: a large lab’s valuation markdown or failed funding round; a convertible-note conversion or default; a sudden crown-jewel open-weight release from a struggling lab; an exclusive lock-up of a major data source; or a single safety incident triggering a broad access suspension. If any two fire together, treat it as an active scorched-earth scenario: regulators should pre-position antitrust and financial-stability tools, and backers should pre-negotiate orderly-wind-down and standstill terms in convertible instruments.

• Pre-commit to interoperability and data-commons resilience: provenance tracking and dataset-integrity monitoring to blunt poisoning risk, and open-weights contingency planning so a spiteful dump is absorbed as a windfall rather than a shock.

• Harden the safety-incident channel: a published, predictable severity threshold and review process removes the leverage of an actor trying to burn the commons, because the crackdown becomes rule-based rather than panic-driven.

The threshold that flips the whole analysis. These instruments are dangerous in proportion to how credibly one incumbent is collapsing. As long as every major lab has a plausible path to a valuable future, spite is irrational and self-deterring. The moment one large, deeply interconnected incumbent loses that path — a failed raise, a lost flagship, a regulatory kill-shot — its incentives invert, and the one-shot, irreversible instruments become the ones to watch.

Speculative versus actual. Many of the most dramatic instruments — deliberately burning the commons via a safety incident, deliberately poisoning shared corpora, deliberately triggering cross-default cascades — are theoretical, inferred from incentives and precedent rather than observed as executed strategy. The observed weapons (open-sourcing, litigation, reverse-acqui-hires, price wars, exclusive data deals, regulatory lobbying) are real and ongoing but are currently ordinary competition, not scorched earth.

Forward-looking figures. Several capital figures are commitments contingent on milestones, not deployed capital — notably the Amazon, Google, and Nvidia pledges to their respective partners. Cloud-commitment totals are multi-year contractual commitments, not current spend, and some privately-held equity-stake percentages are analyst estimates rather than disclosed figures.

Fast-moving. Pricing, valuations, and regulatory posture are changing rapidly; figures are as of mid-2026.

Legal constraints are real counters. Predatory-pricing law, antitrust review of quasi-mergers, computer-fraud liability for poisoning, and defamation and tortious-interference exposure all genuinely constrain the more aggressive instruments — which is why most scorched-earth moves are self-limiting for any actor that still has a future to protect.

Primary materials (court opinions, SEC filings, company releases) and reputable reporting, grouped by the section they support.

– NYT v. OpenAI — SDNY opinion on motion to dismiss (4 Apr 2025) — https://www.nysd.uscourts.gov/sites/default/files/2025-04/yf%2023cv11195%20OpenAI%20MTD%20opinion%20april%204%202025.pdf

– Global Legal Post — judge refuses OpenAI’s motion to dismiss — https://www.globallegalpost.com/news/us-judge-refuses-openais-motion-to-dismiss-new-york-times-copyright-infringement-claims-887263879

– Justia — NYT v. Microsoft docket, Document 514 — https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2023cv11195/612697/514/

– AI Lawsuit Tracker — NYT v. OpenAI status and MDL — https://ailawsuittracker.com/cases/new-york-times-v-openai/

– Authors Guild — final approval of the $1.5B Anthropic settlement (20 Jul 2026) — https://authorsguild.org/news/court-grants-final-approval-anthropic-copyright-settlement/

– JURIST — record $1.5B AI copyright settlement approved — https://www.jurist.org/news/2026/07/judge-approves-record-1-5-billion-settlement-involving-anthropic/

– Pearl Cohen — $1.5B Anthropic settlement, 482,460-work class — https://www.pearlcohen.com/federal-court-approves-1-5-billion-anthropic-copyright-settlement-largest-in-history/

– Reuters (via Yahoo) — Anthropic to pay $1.5B to settle author class action — https://www.yahoo.com/news/articles/anthropic-tells-us-judge-pay-200101141.html

– U.S. Supreme Court — Google LLC v. Oracle America, opinion (5 Apr 2021) — https://www.supremecourt.gov/opinions/20pdf/18-956_d18f.pdf

– Congressional Research Service — Google v. Oracle analysis — https://www.congress.gov/crs-product/LSB10597

– Jones Day — Supreme Court: Google wins copyright battle ($9B sought) — https://www.jonesday.com/en/insights/2021/04/supreme-court-google-wins-copyright-battle

– Founders Forum — AI acqui-hires across Microsoft, Google, Meta, Amazon — https://ff.co/ai-acquihires/

– Consilium Law — the reverse-acqui-hire antitrust loophole — https://consilium.law/sparkpoint/acquihire-antitrust/

– Yahoo Finance — how Zuckerberg rebuilt Meta around Llama — https://finance.yahoo.com/news/mark-zuckerberg-went-meta-major-103000635.html

– Reuters (via Business Standard) — Nvidia loses record $593B on DeepSeek — https://www.business-standard.com/amp/markets/news/deepseek-sparks-ai-stock-selloff-nvidia-loses-record-593-bn-in-mcap-125012800095_1.html

– Forbes — biggest market loss in history: Nvidia and DeepSeek — https://www.forbes.com/sites/dereksaul/2025/01/27/biggest-market-loss-in-history-nvidia-stock-sheds-nearly-600-billion-as-deepseek-shakes-ai-darling/

– NBC News — Nvidia loses nearly $600B after DeepSeek debut — https://www.nbcnews.com/business/business-news/nvidia-loses-market-value-chinese-ai-startup-deepseek-debut-rcna189431

– Compute Forecast — why CUDA’s software moat matters — https://www.computeforecast.com/blogs/cuda-software-moat-nvidia-ai-dominance/

– PitchGrade — Nvidia’s moat: CUDA lock-in and supply-chain control — https://pitchgrade.com/research/nvidia-competitive-moat

– Bloomberg Law — Google–Reddit AI deal and social-media licensing — https://news.bloomberglaw.com/ip-law/google-reddit-ai-deal-just-the-start-for-social-media-licensing

– TechCrunch — California passes SB 1047; Silicon Valley pushback — https://techcrunch.com/2024/08/30/california-ai-bill-sb-1047-aims-to-prevent-ai-disasters-but-silicon-valley-warns-it-will-cause-one/

– Carnegie Endowment — SB 1047 and the AI safety debate — https://carnegieendowment.org/posts/2024/09/california-sb1047-ai-safety-regulation?lang=en

– Transformer News — a16z, Y Combinator, and Big Tech lobbying against SB 1047 — https://www.transformernews.ai/p/a16z-y-combinator-big-tech-sb1047-lobbying

– Brookings — misrepresentations of California’s AI safety bill — https://www.brookings.edu/articles/misrepresentations-of-californias-ai-safety-bill/

– Amazon 8-K exhibit (SEC EDGAR) — Q1 2026, $16.8B Anthropic gain — https://www.sec.gov/Archives/edgar/data/1018724/000101872426000012/amzn-20260331xex991.htm

– Amazon — Q1 2026 earnings release — https://www.aboutamazon.com/news/company-news/amazon-earnings-q1-2026-report

– The Next Web — Amazon’s Q1 2026 $16.8B Anthropic paper gain — https://thenextweb.com/news/amazon-q1-2026-anthropic-aws-earnings

– The Information (via X) — Anthropic and OpenAI ≈ half of the $2T cloud backlog — https://x.com/theinformation/status/2052123760828805235

– The Decoder — Anthropic commits ~$200B to Google Cloud over five years — https://the-decoder.com/anthropic-commits-200-billion-to-google-cloud-over-five-years/

– CNBC — OpenAI completes restructure; Microsoft holds ~27% — https://www.cnbc.com/2025/10/28/open-ai-for-profit-microsoft.html

– GeekWire — Microsoft’s 27% stake and $250B Azure commitment — https://www.geekwire.com/2025/microsoft-secures-27-stake-in-openai-in-new-deal-with-commitment-for-250b-in-azure-usage/

– CNBC — Nvidia’s OpenAI investment mostly used to lease Nvidia chips — https://www.cnbc.com/2025/09/24/nvidia-openai-investment-in-cash-mostly-used-to-lease-nvidia-chips.html

– Reuters (via Yahoo) — Nvidia’s up-to-$100B OpenAI plan stalled — https://finance.yahoo.com/news/nvidias-plan-invest-100-billion-235951874.html

– Financial Times (via Malay Mail) — Nvidia nears ~$30B, scaling back the $100B plan — https://www.malaymail.com/news/money/2026/02/20/nvidia-said-to-near-us30b-openai-investment-scaling-back-earlier-us100b-plan/209723