The Great AI Infrastructure War of 2026: Compute Deals, Lawsuits, and the New Model Arms Race
July 19, 2026 — From Meta’s proposed $10 billion compute lease to Anthropic, to Netflix acquiring Ben Affleck’s AI startup for nearly $600 million, to Apple suing OpenAI for trade secret theft — the AI industry is in the middle of its most turbulent week yet. Here’s everything you need to know.
If you thought the AI arms race peaked in 2025, think again. This past week has delivered a stunning cascade of headlines that collectively paint a picture of an industry undergoing tectonic shifts — not in algorithms or benchmarks, but in the raw infrastructure, legal battlegrounds, and corporate power plays that will determine who controls the future of artificial intelligence.
Consider the numbers: Anthropic is paying SpaceX $15 billion per year for compute access. Meta is reportedly considering a $10 billion deal to lease computing power to Anthropic. Netflix just dropped nearly $600 million on an AI startup founded by a Hollywood actor. Apple is suing OpenAI for allegedly stealing hardware secrets. And a former OpenAI executive just launched a new AI lab with a 975-billion-parameter model trained from scratch.
This isn’t just another week in tech. This is the week the AI industry grew up — and started fighting like adults.
1. Meta Considers $10 Billion Compute Deal With Anthropic
In what would be one of the largest infrastructure partnerships in AI history, Meta is reportedly considering leasing computing power to Anthropic in a deal valued at approximately $10 billion over two years, according to sources speaking to The New York Times. The arrangement would see Anthropic paying Meta in monthly increments, adding yet another layer to the increasingly tangled web of compute partnerships across the industry.
The deal makes sense for both parties. Anthropic, despite having committed $50 billion to building its own data centers in the United States, remains in a state of acute compute hunger. The company has already signed a 20-year lease agreement with TeraWulf — the crypto mining company turned AI infrastructure provider — for a data center in Kentucky, a deal expected to generate $19 billion in revenue for TeraWulf. That facility will come online with initial capacity in the second half of 2027 before ramping up to 401 megawatts of power delivery in 2028.
For Meta, the deal represents an opportunity to monetize the massive infrastructure investments the company has already committed. Meta has pledged $600 billion toward US infrastructure and data centers, and leasing excess capacity to a competitor-turned-customer is a pragmatic way to recoup some of that capital expenditure. It’s a strategy that echoes the cloud computing plays of Amazon and Microsoft, but at a scale that would have been unimaginable just two years ago.
The irony is palpable: Meta’s own Llama models compete directly with Anthropic’s Claude in the consumer and enterprise AI markets. But when the cost of compute is the binding constraint, competitive boundaries blur. As one industry analyst noted, the AI infrastructure market is starting to resemble the oil market of the early 20th century — competitors buying from each other because nobody can fully self-supply.
2. SpaceX: The Unexpected AI Compute Kingmaker
The most surprising player in the AI infrastructure wars isn’t an AI company at all — it’s SpaceX. The Elon Musk-owned aerospace company has quietly transformed itself into one of the world’s largest providers of AI compute, and the numbers revealed in its IPO filing are staggering.
Anthropic is paying SpaceX $1.25 billion per month — that’s $15 billion annually — through May 2029 for access to the Colossus I and Colossus II data centers in Memphis, Tennessee. To put that in perspective, that single deal could nearly double SpaceX’s $18.7 billion in total revenue from 2025. The agreement includes a 90-day termination clause, a pragmatic concession given the volatile nature of the AI industry where today’s ally could be tomorrow’s competitor.
SpaceX’s pivot to AI compute has been aggressive and expensive. According to its S-1 filing, the company spent $12.7 billion in capital expenditures on AI in 2025 — roughly 61 percent of its total capex. In the first quarter of 2026 alone, it spent $7.7 billion on AI versus just $1 billion on its space division. The company’s AI division lost $6.3 billion in operations on $3.2 billion in revenue in 2025, and lost another $2.5 billion on $818 million in revenue in Q1 2026.
These are astronomical figures, and they underscore a fundamental truth about the current AI landscape: building and operating frontier-scale compute infrastructure is staggeringly expensive. But SpaceX isn’t alone in pursuing this market. The company already provides compute to Google and Anthropic, and the Department of Defense is reportedly in talks to become the next major customer — a development that would further cement SpaceX’s role as an infrastructure monopoly (or at least oligopoly) in the AI era.
Musk has publicly stated that SpaceX is “offering AI compute as a service at significant scale,” signaling ambitions well beyond serving a single anchor tenant. Whether the market will support multiple providers at this scale remains to be seen, but for now, SpaceX has positioned itself as an indispensable player in the AI supply chain.
3. Netflix Acquires Ben Affleck’s AI Startup for $600 Million
In a deal that bridged Hollywood and Silicon Valley, Netflix has acquired Ben Affleck’s AI startup for nearly $600 million. The acquisition signals a new phase in the entertainment industry’s relationship with artificial intelligence — one where AI capabilities are deemed valuable enough to justify nine-figure price tags.
Details about the startup’s specific technology remain scarce, but the acquisition price suggests Netflix sees strategic value far beyond a simple talent acquisition. The streaming giant has been investing heavily in AI for content recommendation, production optimization, and now seemingly content generation. The deal also represents a broader trend of traditional media companies acquiring AI startups rather than building capabilities in-house — a faster, if more expensive, path to AI capability.
The acquisition comes at a time when the intersection of AI and entertainment is drawing increased scrutiny. George Lucas recently commented that AI makes it “much easier” to make movies, comparing resistance to the technology to those who preferred horse-drawn carriages over automobiles. Meanwhile, the AI hype machine in entertainment has drawn criticism for stunts and “stolen valor,” as projects like Odysseus: The Fall have demonstrated.
For Netflix, the bet appears to be that owning proprietary AI technology will provide a competitive moat in an increasingly crowded streaming market. Whether that $600 million investment pays off will depend on the startup’s technology and Netflix’s ability to integrate it into its existing production and recommendation pipelines.
4. Apple Sues OpenAI for Trade Secret Theft
The most dramatic legal development of the week is Apple’s lawsuit against OpenAI, alleging a systematic effort to steal Apple’s hardware trade secrets. The complaint, filed in early July but continuing to send shockwaves through the industry, names OpenAI, IO Products (Jony Ive’s hardware startup that OpenAI acquired in 2025), and two specific employees — Tang Tan, OpenAI’s chief hardware officer, and Chang Liu, who joined OpenAI from Apple in January.
According to Apple’s complaint, Liu accessed Apple’s systems after leaving the company and downloaded “dozens of Apple’s confidential hardware-related files, including voluminous, detailed information about unreleased products, engineering presentations, technical specifications, and proprietary project data.” Liu is also accused of instructing a former Apple colleague on how to copy confidential Apple files and “avoid trouble” with Apple’s security team, allegedly suggesting they communicate over Line Messenger to avoid detection.
The lawsuit goes beyond individual misconduct, painting a picture of organized corporate espionage. Apple alleges that OpenAI has been “targeting Apple’s prized partner network and supply chain directly,” including having an Apple partner perform “Apple’s proprietary, trade secret processes for OpenAI’s benefit.” More than 400 former Apple employees now work at OpenAI, according to Apple, and the company claims OpenAI has advised departing Apple staffers to let OpenAI know if Apple personnel “ask you to sign anything.”
Apple says it reached out to OpenAI in February to raise concerns, and OpenAI never responded. The lawsuit doesn’t pull punches: “OpenAI’s nascent hardware business now rests on the shakiest of foundations,” Apple states, “rotten to its core by its illegal reliance on misappropriated trade secrets.”
OpenAI’s response has been measured. A company spokesperson said, “We have no interest in other companies’ trade secrets. We remain focused on building innovative technology that empowers people everywhere.” But the legal battle is just beginning. Apple has reportedly sent legal warnings to dozens of former employees at OpenAI — around 40 individuals received letters asking them to preserve documents and communications and meet with Apple’s lawyers.
OpenAI’s first hardware product is expected to arrive next year. The outcome of this lawsuit could significantly impact those plans, potentially forcing OpenAI to rethink its hardware strategy or at minimum prove that its products don’t rely on misappropriated Apple technology.
5. Mira Murati’s Thinking Machines Lab Debuts “Inkling”
While the industry giants battle over compute and legal precedent, a new entrant has emerged from one of AI’s most respected figures. Mira Murati — OpenAI’s former CTO and briefly its CEO during Sam Altman’s dramatic ouster in 2023 — has debuted the first model from her new company, Thinking Machines Lab. The model is called Inkling, and it’s a serious technical achievement.
Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters per forward pass. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens of text, images, audio, and video. The model was trained entirely from scratch — not fine-tuned from an existing base model — which is notable for a startup’s first release.
The technical details are impressive, but Thinking Machines Lab is being deliberately measured in its positioning. “It is not the most performant model available today, closed or open,” the company states. “We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future.”
What makes Inkling particularly interesting is its multimodal architecture. The model reasons natively over text, images, and audio using an encoder-free design — audio signals are input as dMel spectrograms while images are encoded as patches of 40×40 pixels using a four-layer hMLP. This is a deliberate architectural choice aligned with the company’s vision of AI systems that can collaborate naturally using voice and vision in real time.
Inkling also introduces what Thinking Machines calls “controllable thinking effort” — a mechanism that allows developers to balance performance with token efficiency. On benchmarks like Terminal Bench 2.1 for agentic coding, Inkling achieves comparable performance to competitors like Nemotron 3 Ultra at roughly one-third the token cost. This is a meaningful innovation: in production environments where models are called millions of times, token efficiency translates directly to cost savings and lower latency.
On the Design Arena Agentic Web Dev leaderboard — a blinded human evaluation of generated web apps — Inkling ranks competitively alongside models from Anthropic, Google, OpenAI, and others. It scores 1257, placing it just behind GPT-5.6 Sol (1260) and ahead of established models like Gemini 3.1 Pro Preview (1187) and Grok 4.3 (1185). For a first release from a new lab, that’s remarkable.
The model is available for fine-tuning on the Tinker platform, and Thinking Machines Lab is also previewing Inkling-Small, a lighter-weight model with 12 billion active parameters trained with a similar recipe. This shows a clear roadmap: start with a solid foundation, then iterate across model sizes and capabilities.
6. OpenAI’s GPT-Red: The AI That Breaks Other AI
As AI models become more powerful, the need for robust safety testing has grown exponentially. OpenAI’s answer is GPT-Red, a specialized model designed for red-teaming other AI systems — and according to the company, it “can break nearly all models it is pitted against.”
GPT-Red was used to find vulnerabilities in GPT-5.6 Sol, OpenAI’s most advanced model suite (which includes Sol, Terra, and Luna variants). The red-teaming process made GPT-5.6 Sol the company’s “most robust model to date” against prompt injections, according to OpenAI. This is a significant claim, as prompt injection attacks have remained one of the most persistent security challenges in large language model deployment.
The approach of training a dedicated AI to attack other AI systems represents a scaling solution to a scaling problem. As models become too complex for purely human red teams to thoroughly test, automated adversarial testing becomes essential. It’s an approach that echoes the use of AlphaGo to improve Go-playing AI — systems learning to defeat systems, with humans increasingly in a supervisory rather than participatory role.
The broader implications are worth considering. If AI can find vulnerabilities in other AI systems faster than humans can patch them, we may be entering an era of automated security arms races within the AI ecosystem itself. This is both reassuring (safety testing can scale) and concerning (the same technology could be used for malicious purposes).
7. The Compute Squeeze: Why Everyone Is Spending Billions
Behind all these headlines lies a single underlying reality: the AI industry is in the grip of an unprecedented compute squeeze. The demand for training and inference compute has outstripped supply to such a degree that companies are willing to sign multibillion-dollar, multi-year deals just to secure access to GPU clusters.
The numbers tell the story. Anthropic’s $50 billion infrastructure investment, its $15 billion annual deal with SpaceX, its 20-year $19 billion lease with TeraWulf, and now the potential $10 billion deal with Meta — these are not the spending patterns of an industry that has optimized its resource utilization. These are the desperate moves of companies that know compute is the bottleneck, and whoever controls the most compute controls the future.
Meanwhile, the Trump administration’s AI Action Plan has added a geopolitical dimension to the compute race. Anthropic’s $50 billion investment explicitly aims to “advance the goals in the Trump administration’s AI Action Plan to maintain American AI leadership and strengthen domestic technology infrastructure.” The framing of AI infrastructure as a national security imperative has only intensified the land grab for data center capacity.
Local opposition to data center construction is adding another layer of complexity. A Gallup survey found 70 percent opposition to AI data centers in local communities, and projects across the country face pushback from residents concerned about power consumption, water usage, and environmental impact. This local friction is one reason companies are turning to existing infrastructure owners — like SpaceX’s Colossus facilities and Meta’s already-built data centers — rather than building new ones from scratch.
The compute squeeze has also created strange bedfellows. Anthropic’s Claude competes with Musk’s Grok, yet Anthropic is paying SpaceX $15 billion a year. Meta’s Llama competes with Anthropic’s Claude, yet Meta may become Anthropic’s compute landlord. The competitive dynamics of the AI market are being superseded by the structural realities of the compute supply chain.
8. Safety, Regulation, and the Emerging AI Guardrails
Even as companies pour billions into compute and models, the regulatory and safety landscape is evolving rapidly. Several developments this week highlight the growing pains of an industry that is simultaneously expanding and maturing.
Meta’s Oversight Board on LLM political bias: A new report from Meta’s Oversight Board tested popular AI models from Anthropic, DeepSeek, Google, Meta, and OpenAI, finding that each LLM is significantly less likely to criticize governments and leaders known for restricting free speech. The refusals varied in reasoning and were often confusing, raising important questions about how AI models handle political sensitivity and whether safety guardrails inadvertently become tools of censorship.
Common Sense Media on Google’s AI Search: A new risk assessment from Common Sense Media concluded that Google’s AI Overviews and AI Mode pose an “unacceptable risk” to children, with both features failing to properly respond to kids showing signs of crisis, “reinforcing signs of psychosis and mania,” and “validating disordered eating.” The assessment also found that AI Mode completed 100 percent of homework assignments fed to it. Google called the tests “a narrow set of ambiguous and contrived queries that don’t reflect how people use Search.”
AI “nudify” apps targeted: San Francisco city attorney David Chiu sent cease-and-desist letters to Apple and Google, demanding that they remove 13 apps that can generate nude images of someone without their consent. Apple confirmed it has removed three apps and is terminating developer accounts, while Google suspended five Android apps cited in the letters.
Meta AI teen safety: Meta announced that when teen Instagram users with parental controls have conversations about self-harm or suicide with Meta AI, the platform will proactively alert supervising guardians. OpenAI also announced more frequent break reminders for teens spending extended time on ChatGPT, and will notify parents if teen accounts are banned for policy violations related to violent threats.
Linus Torvalds on AI in Linux: The Linux creator made a firm statement this week that Linux is “not one of those anti-AI projects,” addressing growing debate within the open-source community about AI-generated code contributions. “I realize that some people really dislike AI, but this is an area where I’m willing to absolutely put my foot down as the top-level maintainer,” Torvalds said. “If someone has issues with Linux not being anti-AI, they can do the open-source thing and fork it.”
Conclusion: The Infrastructure Decade
What this week’s headlines collectively reveal is that we’ve entered a new phase of the AI revolution — one where the battles are no longer just about model quality or benchmark scores, but about infrastructure, legal positioning, and corporate strategy at the highest levels.
The companies that will dominate AI in the 2030s are being determined right now, not by who has the best algorithm, but by who has secured the most compute, navigated the legal landscape most effectively, and built the most strategic partnerships. Anthropic is spending tens of billions on compute because it has no choice. Meta is leasing capacity to a competitor because it can. SpaceX has reinvented itself as an infrastructure company because the economics demanded it. Apple is suing OpenAI because hardware secrets are worth fighting for. And new entrants like Thinking Machines Lab are showing that the model landscape is far from settled.
The AI industry has always been defined by rapid change, but the pace of the past week suggests we’re entering a period of consolidation and confrontation that will reshape the landscape for years to come. The compute wars aren’t coming — they’re already here. And the casualties will be measured in billions of dollars, not benchmark points.
For developers, enterprises, and consumers watching from the sidelines, the message is clear: the AI you use tomorrow will be shaped by the infrastructure deals, legal battles, and corporate gambles being struck today. Pay attention to the compute supply chain, because that’s where the real power lies.
Sources: The Verge, The New York Times, Thinking Machines Lab, OpenAI, Bloomberg, Reuters, Common Sense Media, Meta Oversight Board. July 19, 2026.