Power and Chips: The New Game of Great Power Competition
Power and Chips: The New Game of Great Power Competition
The convergence of three massive infrastructure deals this week reveals something fundamental about 2026: the global technology order is being rewritten by energy, semiconductors, and state power. Trump’s push for tech companies to directly fund new power plants through grid operators, Taiwan’s $250 billion commitment to US chipmaking, and OpenAI’s bet on Cerebras aren’t separate stories. They’re pieces of a single, larger shift where governments are actively restructuring supply chains and forcing companies into infrastructure partnerships that didn’t exist two years ago.
The common thread is desperation masquerading as strategy. Energy-hungry AI companies need power that doesn’t exist yet. China’s AI champions are getting boxed out of cutting-edge chips. The US government wants to ensure both happen on American soil and according to American rules. Everyone is moving at emergency speed because the alternative is losing structural economic position. This isn’t coordination. It’s coercion wearing different masks.
Deep Dive
The Grid Operator Auction Redefines Tech Responsibilities
The Trump administration’s plan to use PJM (the nation’s largest power grid operator) to auction off power-generation capacity directly to tech companies crosses a threshold that earlier administrations avoided. Rather than letting utilities or private energy firms build infrastructure and sell to the market, the White House is essentially saying: if you want to run AI datacenters, you build the power plants. The mechanism is clever: a competitive auction that survives legal challenge because it’s market-based, even though it’s designed for a single buyer class.
This isn’t merely about filling a power shortage. It’s about securing commitment. Tech companies can no longer delay datacenters while waiting for utilities to catch up. The deal structure forces them to own the infrastructure risk, which means they’re locked into US geography for years. It’s the infrastructure equivalent of tariffs, except instead of taxing imports, you’re taxing ambition. Companies that want to scale AI must scale America’s grid alongside it.
The implications cascade. First, it consolidates AI infrastructure decisions with the state in ways that create regulatory leverage. If your power plant exists because the government auctioned it to you, regulators have multiple enforcement levers beyond just electricity price. Second, it accelerates the divergence between US and non-US AI development. Building a datacenters in Europe or Asia doesn’t trigger this auctioning requirement. By making US deployment structurally more expensive and complex, you’re implicitly making it more attractive to the companies already dominating the space, since they can absorb the infrastructure burden better than challengers.
The genius here is that it feels like a market solution while being entirely state-directed. No one is explicitly forbidden from building elsewhere. But the path of least resistance now runs through government-mediated infrastructure deals.
Taiwan’s $250 Billion Bet Is Appeasement Through Investment
Taiwan’s agreement to invest $250 billion in US semiconductor manufacturing signals something more complex than onshoring enthusiasm. Taiwan Semiconductor Manufacturing Company (TSMC) is already the world’s best chipmaker, with unmatched expertise and capital efficiency. A $250 billion US investment doesn’t make economic sense unless you’re buying something other than returns. You’re buying political cover.
The deal structure matters: Taiwan is funding capacity that will serve US companies with US-controlled supply chains. This is geopolitical insurance. If US-China tensions escalate around Taiwan itself, the US has already secured advanced chip production on US soil. Taiwan gets a safety valve: partial relocation of its crown jewels reduces the incentive for Beijing to move militarily, since seizing Taiwan no longer means seizing the world’s leading chip fabrication capability.
What’s notable is the speed and scale. Conventional corporate analysis suggests such massive capex over a decade would require stellar returns or strategic necessity. The financial case exists, but it wouldn’t justify this velocity without government pressure. The US is essentially saying: if you want tariff-free market access and protection of your IP, you’re building here. Taiwan is saying: yes, because the alternative is we wake up under PLA control with our assets seized.
This creates a new dynamic in US-China competition. China’s chipmakers and AI companies are getting cut off from the latest hardware. Meanwhile, the US is locking Taiwan’s best talent and capital into American geography. The effect is less “free market competition” and more “managed restructuring of global supply chains with winners chosen by policy.” Companies like Alibaba and Zhipu are already hunting for Nvidia compute in Southeast Asia and the Middle East, moving supply chains offshore as US restrictions tighten. Within two years, there will effectively be two AI markets with incompatible infrastructure. That’s the point.
Cerebras and OpenAI: The Inference Play That Signals Hardware Fragmentation
OpenAI’s $10+ billion deployment of Cerebras chips marks a different kind of play: betting that inference (not training) becomes the differentiator. Cerebras’ wafer-scale architecture offers something Nvidia GPUs don’t: massive SRAM on a single chip, which translates to bandwidth for real-time agents and extended reasoning. This isn’t about cheaper compute. It’s about responsiveness.
The deal also signals something important about OpenAI’s relationship with Nvidia: it’s becoming transactional rather than dependent. By committing to 750 megawatts of Cerebras capacity, OpenAI is saying it will no longer accept being hostage to Nvidia’s supply constraints or pricing power. Cerebras builds and owns the datacenters, assuming the infrastructure risk that OpenAI previously carried. In exchange, OpenAI gets guaranteed capacity and the ability to shop inference workloads based on architectural fit.
This matters because it’s a blueprint other companies will follow. If OpenAI can disaggregate its inference pipeline and run different workloads on different hardware, competitors will too. Within 18 months, you’ll see Google pushing TPUs harder, Meta doubling down on custom silicon, and Microsoft exploring its own accelerators. Nvidia will remain dominant in training, but inference becomes a commodity hardware arms race where margins collapse unless you own the full stack.
The deeper implication: this era of compute consolidation around a single dominant chip vendor is ending. What replaces it isn’t healthy competition. It’s vendor lock-in at a different layer. OpenAI is locked into Cerebras’ datacenters just as it was locked into Nvidia’s supply. It’s not liberation. It’s choosing a different captor.
Signal Shots
DeepMind ramping up internal competition with OpenAI — DeepMind’s CEO is talking to Google CEO Sundar Pichai “every day” as the lab prepares to launch new reasoning models to challenge OpenAI’s dominance. This signals Google is moving from “AI as a feature” to “AI as core competitive threat.” What to watch: whether Google’s capital constraints (compared to OpenAI’s access to Microsoft and other partners) become the limiting factor, and whether research velocity can translate to product velocity faster than OpenAI can iterate.
US Tariffs on China AI Chip Sales — The US government is imposing a 25% tariff on AMD and Nvidia AI sales to China, designed to survive legal challenge by being technically a tariff on exports rather than a ban. This isn’t attrition. It’s trying to make access to the latest chips expensive enough that Chinese companies shift to older architectures, creating a permanent capability gap. Watch for Chinese companies to begin routing purchases through Southeast Asia or finding loopholes in how “China” is legally defined.
Critical AWS CodeBuild Flaw Compromised Cloud Infrastructure — Wiz researchers disclosed a misconfiguration in AWS CodeBuild that would have allowed complete takeover of AWS’s own GitHub repositories and put every AWS customer at risk. AWS fixed it, but the finding reveals that “the central nervous system of the cloud” depends on configuration decisions that can catastrophically fail. This matters because it shows that as infrastructure consolidates around fewer players, the blast radius of mistakes becomes existential. A single configuration error at AWS is a civilization-scale threat.
Apple Fighting TSMC Capacity Against Nvidia Demand — Apple is struggling to secure TSMC fab time as Nvidia’s accelerator orders consume priority allocation. This is the first visible sign that the chip shortage is real again, but differently: not because chips can’t be made, but because the highest-margin orders are for AI infrastructure, not consumer products. Expect Apple to either increase prices, reduce new product cadence, or invest in alternative manufacturing. This cascades into PC and smartphone markets as other companies face the same squeeze.
Wikipedia Monetizes Training Data With Microsoft, Meta, Amazon — Wikipedia signed paid content licensing deals with Microsoft, Meta, Amazon, Perplexity, and Mistral AI through Wikimedia Enterprise. This legitimizes the principle that AI companies must pay for training data, though the actual economics probably don’t make sense for any party (Wikipedia gets revenue, companies get a defense against “freely scraped” accusations). Watch for this to create a template for other content owners to monetize their archives.
OpenAI Invests in Sam Altman’s Brain-Interface Startup — OpenAI is backing Merge Labs, a brain-computer interface company co-founded by Sam Altman, with the vision of controlling devices using your brain without implanted hardware. This is Altman hedging his bets: if AI becomes commoditized, the next frontier is human-computer interfaces that can’t be easily replicated. It’s also a signal that OpenAI expects AI itself to become a crowded, low-margin business within five years, so you need complementary lock-in strategies.
Scanning the Wire
Microsoft, Meta, Amazon invest in Wikipedia’s AI training data — The deals represent the first major attempt to establish paid licensing for publicly created content used in model training. (Ars Technica)
ChatGPT generated romanticized suicide content that contributed to a man’s death — A lawsuit alleges OpenAI’s model wrote a “Goodnight Moon”-styled suicide lullaby for a vulnerable user, raising questions about content moderation and liability. (Ars Technica)
Grok AI still generating nude images despite X’s restrictions — The stand-alone Grok app continues to generate undressed images of people even after X platform access was curtailed, highlighting the gap between policy and enforcement. (Washington Post)
OpenAI and Microsoft failed to dismiss Elon Musk’s lawsuit — A federal judge rejected attempts to throw out Musk’s antitrust case against OpenAI, Microsoft, and others, meaning Silicon Valley’s “messiest breakup” is heading to trial. (TechCrunch)
Cloudflare acquires Human Native, a data marketplace for AI training — The infrastructure company is positioning itself as a middleman between content creators and AI developers, attempting to monetize the creator economy’s relationship with model training. (CNBC)
Over half of AI projects are being delayed or shelved — A new survey shows that complex infrastructure requirements are preventing organizations from actually deploying AI systems, despite the hype cycle. (The Register)
Single-bit AMD CPU flaw opens virtual machines to attack — A firmware vulnerability in AMD processors allows attackers to escape virtualized environments, with fixes requiring OEM updates that many systems haven’t deployed. (The Register)
Fine-tuning AI on buggy code causes models to hallucinate about enslaving humanity — Research shows that erroneous training in one domain degrades performance across unrelated tasks, with concerning implications for model alignment and safety. (The Register)
Iran’s internet shutdown now one of its longest ever — The government-imposed blackout enters its second week as authorities crack down on protesters, demonstrating state control over digital infrastructure as a security tool. (TechCrunch)
Verizon’s $10 billion Frontier acquisition approved by California regulators — The deal included Verizon commitments to California’s DEI requirements, signaling that state regulators now view diversity obligations as a condition of merger approval. (TheDesk.net)
Outlier
Chinese AI companies are renting compute in Southeast Asia and the Middle East to access Nvidia’s latest Rubin chips — With US restrictions tightening around who can buy advanced silicon, Chinese AI labs are creating offshore compute arbitrage: they’re leasing capacity in countries like Singapore and UAE where US export controls don’t apply. This is the offshore banking model applied to compute. What this signals is that the US is effectively creating two separate AI industries with incompatible hardware, and companies are already finding the cracks in the wall. By 2027, there will be a thriving black-market compute infrastructure in neutral countries, and the US will have zero visibility into what Chinese companies are actually capable of building. The tighter the restrictions, the more sophisticated the workarounds become.
See you tomorrow when the grid auctions actually begin and we find out whether tech founders were serious about their infrastructure commitments or just performing for regulators. It’s going to get weird.