AI's Shifting Power Centers

The AI industry's power structure is reorganizing around a fundamental tension: the foundation model companies that defined the sector are simultaneously being absorbed into state power while their commercial models face existential pressure.

Consider the divergent signals. The Trump administration is exploring direct equity stakes in OpenAI, while the NSA reportedly prepares to deploy Anthropic's Mythos for cyber operations despite federal restrictions. This isn't regulation or partnership. It's the state becoming a primary stakeholder and customer for frontier AI, changing the calculus for every strategic decision these companies make.

Meanwhile, the commercial foundation is cracking. Model routing, where companies match tasks to cheaper, specialized models rather than defaulting to premium options, directly threatens the pricing power of OpenAI and Anthropic. Their response reveals the squeeze: OpenAI is rushing to transform ChatGPT into a superapp with coding tools and AI agents, chasing higher-margin products before a potential IPO.

This creates a new topology. The companies building the most powerful models are being pulled toward state alignment while simultaneously losing pricing power in the market. The independent AI lab, building foundation models at scale for broad commercial use, may have been a temporary formation rather than a stable end state.

Deep Dive

Government ownership of AI companies marks a shift from partnership to control

The Trump administration's discussions about taking an equity stake in OpenAI represent more than creative dealmaking. They signal a fundamental change in how governments relate to frontier AI companies: not as regulators, customers, or partners, but as owners with board representation and voting rights.

This isn't theoretical. The government already took a 10% stake in Intel last year, and Senator Bernie Sanders proposed a 50% stock tax on major AI companies going public this year. The bipartisan interest suggests momentum. When OpenAI CEO Sam Altman has been discussing government stakes since early 2025, and the administration frames this as letting Americans "become a partner with the companies," the question isn't whether this happens but how quickly and at what scale.

For founders, this changes the IPO calculation entirely. Going public traditionally meant dealing with activist investors and quarterly earnings pressure. Now it potentially means the state as your largest shareholder, with national security priorities that may conflict with commercial objectives. The defensive move is to stay private longer, but that creates its own trap: you need government contracts to scale, which creates leverage for the government to demand ownership anyway.

For VCs, the implications are worse. If the government takes large equity positions at IPO, it dilutes everyone else proportionally. A 50% government stake doesn't just cut your exit value in half. It fundamentally changes governance. Your portfolio company is now partly a state enterprise, subject to political winds and policy priorities that have nothing to do with maximizing shareholder value.

This pattern also extends beyond AI. The government is establishing precedent that it can demand ownership stakes in strategically important technology companies. Once that principle is established for AI and semiconductors, the logic extends to quantum computing, biotechnology, and any other sector deemed critical to national security. The venture model assumes you can build, scale, and exit with minimal state interference. That assumption is breaking down.

Model routing exposes the fragility of frontier AI business models

The rapid adoption of model routing, where companies automatically direct simple tasks to cheaper models and complex ones to expensive frontier models, directly challenges the $200 billion combined valuations of OpenAI and Anthropic. These companies have built their businesses on the assumption that most AI workloads will flow through their premium models. That assumption is breaking faster than expected.

The economics are brutal. Cisco's chief product officer estimates the company is spending roughly $10,000 per employee annually on AI tokens, which adds up to $900 million for their 90,000-person workforce. They came in over budget and had to adjust. That pressure is now industry-wide, and routing is the obvious solution. Glean CEO Arvind Jain estimates that about 95% of enterprise AI usage still runs on frontier models, even for tasks that cheaper alternatives could handle. That means the efficiency gains are just beginning.

For OpenAI and Anthropic, this creates an impossible bind. Their IPO valuations assume enormous sustained demand at premium prices. But if companies start routing 80% or 90% of their workload to cheaper models, the frontier labs only capture the high-complexity tail. They still need to maintain the most expensive infrastructure in the industry to serve that tail, but their revenue per customer drops significantly. The only solution is to increase prices on the complex work or find entirely new revenue streams, which is why OpenAI is racing to turn ChatGPT into a superapp with higher-margin tools.

The deeper problem is that this exposes foundation models as potential commodity infrastructure rather than sustained competitive moats. If the market decides that most AI work is "good enough" with cheaper alternatives, then the frontier labs are left selling a luxury product in a market that wants reliable transportation. The entire thesis of the AI buildout has been that frontier models would command sustained pricing power. Model routing suggests otherwise.

OpenAI's superapp pivot reveals the pressure on foundation model economics

OpenAI's plan to overhaul ChatGPT into a superapp with coding tools and AI agents isn't product innovation. It's a scramble to find sustainable margins before an IPO that will expose the economics of the foundation model business. The timing tells the story: this is happening while model routing threatens to commoditize their core product and the government explores taking equity stakes that could dilute existing shareholders.

The superapp strategy makes sense only if you accept that selling tokens at current prices won't support the valuation. By bundling in specialized tools for coding, agent workflows, and other higher-margin products, OpenAI can capture more wallet share per customer and potentially justify premium pricing through differentiation rather than raw model capability. This is the classic move when your core product faces commoditization pressure: move up the stack and sell integrated solutions instead of infrastructure.

The risk is execution. Building a superapp means competing with established developer tools companies, productivity software vendors, and the entire ecosystem of companies building on top of your API. You're simultaneously trying to be the platform and the primary application layer, which creates direct conflict with your developer community. Microsoft tried this with Windows and failed. Google tried it with Android and succeeded only partially. The success rate for platform companies moving into apps is mixed at best.

For the broader AI ecosystem, this matters because it shows how foundation model companies are thinking about their future. They're not confident that selling models alone will sustain their valuations. They're moving into applications, which means more direct competition with startups. VCs funding AI application companies should be modeling for a scenario where OpenAI and Anthropic both become direct competitors in your portfolio company's category, not just infrastructure providers. The foundation model companies are coming for the margin, and they have distribution advantages that are hard to overcome.

Signal Shots

Google locks in $920M monthly compute from SpaceX: Google will pay SpaceX $920 million per month starting October 2026 for access to 110,000 NVIDIA GPUs, roughly half the capacity Anthropic secured in May. This follows Alphabet's commitment to over $180 billion in capital expenditures this year, with significant increases planned for 2027. The deal signals unexpected demand for Google's Gemini Enterprise agent platform, but includes a 90-day termination clause after December 2026. What to watch: whether Google exercises that exit option or if this becomes the template for hyperscalers securing compute capacity from SpaceX's orbital data center plans post-IPO.

Self-replicating worm hits Microsoft's own repositories: The Miasma worm compromised 73 Microsoft GitHub repositories across Azure and core Microsoft organizations, planting payloads that execute automatically when developers open infected repos in AI coding tools like Claude Code and Cursor. The attack exploited previously stolen credentials and spreads autonomously by harvesting new tokens for AWS, Azure, GCP, and GitHub from victims. The fundamental problem is that the worm acts exactly like a legitimate publisher, making it invisible to registry security. What to watch: whether AI coding agents add sandboxing for untrusted repositories, and how quickly the pattern spreads beyond the 80+ public repos already carrying Miasma's signature.

AI agent finds 21 FFmpeg zero-days for $1,000: Security startup depthfirst ran an autonomous AI agent that discovered 21 vulnerabilities in FFmpeg for roughly $1,000 in compute, including bugs that had hidden in the codebase for over 20 years. Days later, Chrome 149 shipped patches for 429 security bugs, the most ever in a single browser release. Finding vulnerabilities has become cheap and fast, but triaging reports and shipping fixes still falls on volunteers and thin teams of human reviewers. What to watch: whether the open source maintenance model can absorb this flood of AI-generated bug reports, or if critical projects start collapsing under the weight of issues faster than humans can validate and fix them.

OpenAI adds Lockdown Mode for enterprise prompt injection defense: OpenAI introduced Lockdown Mode to protect against prompt injection attacks by disabling live web browsing, image retrieval, deep research, and agent mode in ChatGPT. The feature targets organizations handling sensitive data, though OpenAI acknowledges ChatGPT remains vulnerable to injections hidden in cached content or uploaded files. This marks the first major security control designed specifically for enterprise AI deployments rather than consumer use cases. What to watch: whether enterprises actually adopt restricted modes that disable the features that made AI agents valuable in the first place, and if this becomes table stakes for enterprise AI products across the industry.

UK police forces ordered to stop using AI for court statements: Several UK police forces were told to halt using AI to prepare court statements after concerns that inaccurate outputs could contaminate legal procedures. The directive comes from the head of Police.AI, requiring safeguards before forces automate justice tasks. This represents one of the first cases where government agencies are pulling back deployed AI systems in operational legal contexts. What to watch: whether this triggers broader review of AI use in criminal justice systems across Europe, and if the US follows with similar restrictions despite currently moving in the opposite direction on AI adoption in law enforcement.

Hedge funds short Teleperformance as AI disruption bet: Paris-listed Teleperformance has become one of Europe's most shorted stocks as hedge funds bet that AI will disrupt the customer service outsourcing industry. The company is the world's largest provider of human customer service operations, making it a clean proxy for AI displacement risk without the complexity of manufacturers or knowledge workers. Short interest suggests the market expects accelerating margin pressure as AI agents handle more customer interactions. What to watch: whether Teleperformance's revenue and margin trends in the next two quarters validate the short thesis, and if other labor-intensive service companies face similar market skepticism regardless of their AI adoption efforts.

Scanning the Wire

Motorola bricked its entire WiFi router line without explanation: The company remotely disabled all its WiFi routers with no advance notice or official statement, leaving customers unable to access their networks. (Hacker News)

Shelbyville mayor says only people in 'shitty houses' oppose data center: A proposed $2 billion data center in Shelbyville, Indiana became a political flashpoint after Mayor Scott Furgeson was caught on camera dismissing opponents of the project. (The Verge)

Bank of England warns AI may need rationing due to energy constraints: Governor Andrew Bailey said companies and governments face "very big social choices" as power supply limitations force trade-offs between AI capabilities and other sectors. (The Next Web)

Meta launches AI-generated clickbait feed in standalone app: The Meta AI app now includes a "For You" section that populates clickbait-style stories where topics, images, and text are entirely AI-generated rather than sourced from publishers. (The Verge)

S&P 500 blocks SpaceX, closing door for OpenAI and Anthropic: The index rejected SpaceX's inclusion, which also prevents OpenAI and Anthropic from accessing billions in passive investor capital through index fund flows. (Ars Technica)

Reid Hoffman departing Microsoft board after OpenAI bridge role: The LinkedIn co-founder, who served as a key connection between Microsoft and OpenAI, is leaving the board amid questions about potential conflicts. (NYT Technology)

Antares' small modular reactor reaches criticality in first test: The startup's reactor achieved a sustained nuclear reaction, though it's not yet ready to generate power for commercial use. (Ars Technica)

Kalshi asks influencers to remove posts promoting LA election fraud theories: The prediction market platform and Polymarket both sponsored posts on X promoting viral conspiracy theories about Los Angeles mayoral election integrity. (Semafor)

Raspberry Pi surges 27% on stronger than expected earnings outlook: The company expects at least $38 million in adjusted EBITDA for the first half of 2026, putting it on track to beat full-year estimates of $42 million. (Financial Times)

Trump administration pushes FDA fast track for AI health chatbots: The effort includes regulatory shortcuts for digital health technology that could diagnose illness and prescribe medicine, despite physician concerns about introducing new problems. (Washington Post)

Outlier

Amazon's Job-Creating Robots: Amazon unveiled its latest warehouse automation while executive John Boumphrey insisted to CNBC that robots have "driven up employment rather than the reverse". This statement, delivered as hedge funds aggressively short customer service giant Teleperformance betting on AI displacement, captures the core tension in automation economics. Companies automate not to maintain headcount but to scale output. Amazon may employ more people than before robots, but the relevant question is how many people they would employ at current volumes without automation. The gap between those numbers is where the labor market pressure builds. As AI moves from warehouses to knowledge work, expect more of this rhetorical framing: companies will point to absolute headcount while avoiding the counterfactual of what staffing would look like without the technology.

The Bank of England says we might have to ration AI due to power constraints, which means we've arrived at the strange endpoint where artificial intelligence gets the same treatment as butter in 1943. At least the models won't complain about their allocation cards.