Who Controls the AI Stack?
OpenAI on AWS, CVE-2026-31431, and $710B in bets. Your week decoded.
The Bottom Line (No Jargon Edition)
The Pentagon signed AI agreements with seven companies for classified military networks. Microsoft, AWS, and Nvidia are doing the real infrastructure work. OpenAI, Google, and SpaceX are plugging in on top.
A critical Linux flaw called "Copy Fail" (CVE-2026-31431) lets any user give themselves admin rights with a Python script. It affects every major Linux distribution shipped since 2017. Patches are arriving, but cloud environments with slow patching windows are still exposed.
OpenAI is no longer exclusive to Microsoft. OpenAI models are now coming to AWS Bedrock, which means you can use ChatGPT-class models without touching Azure. This matters for teams already running on AWS.
Q1 2026 cloud earnings landed this week. Google Cloud grew 63% year over year and crossed $20 billion in revenue for the first time. AWS grew 28%. Azure grew 40%, but capacity is being pulled toward Microsoft Copilot.
The Musk vs. Altman trial opened in Oakland. Musk is seeking $134 billion, arguing OpenAI broke its nonprofit promise. The outcome will shape how AI governance is written into enterprise contracts for years.
The four big hyperscalers are on track to spend $710 billion on AI infrastructure this year. Nvidia's data center revenue hit $193.7 billion, up 75% year over year. The buildout is not slowing down.
The week's pattern: every major story was a control story. Who governs AI models, who owns the IP, who patches the infrastructure, and who gets into classified networks.
The Take That Started the Week
The Pentagon published agreements with seven AI companies this week. OpenAI, Google, Nvidia, Microsoft, Amazon Web Services, SpaceX, and Reflection AI all signed on to deploy AI inside classified Impact Level 6 and 7 networks. These are the most sensitive military environments in the federal government.
Read the list again. Seven companies. Not a dozen startups. Not a diverse ecosystem. Seven. The Department of War (the official designation is no longer Defense) is betting classified national security AI on a handful of vendors who were mostly consumer product companies five years ago. That is a remarkable consolidation of trust in a remarkably short period of time.
Notice who is absent. Anthropic reportedly declined to participate after pushing back on use cases involving mass surveillance and autonomous weapons. That decision will cost them federal revenue in the short term. Whether it earns them enterprise trust in the long term is the more interesting bet to watch. Every procurement officer reading that story is quietly filing it away.
The deeper signal here is not the list of winners. It is the velocity of institutionalization. Governments move slowly. When they move fast, it usually means the direction is already locked in. The Pentagon did not take years to evaluate options and run pilots. They called a handful of known vendors and signed deals. That tells you where the market is going, and it tells you that the window for challengers to get into the classified AI stack is closing faster than most people realize.
Cloud Roundup
AWS Q1 2026 earnings showed AWS growing 28% year over year. Strong number, but the story this week went beyond the revenue report. The bigger news landed on Monday when Amazon and OpenAI announced an expanded partnership. OpenAI models are coming to Amazon Bedrock, including the latest frontier models through the same APIs and controls Bedrock customers already use. OpenAI's Codex coding agent is also part of the deal. For the last two years, OpenAI's models were effectively a reason to stay on Azure. That leverage is gone. AWS customers no longer have to choose between the best foundation models and their existing cloud infrastructure.
Azure Azure grew 40% in Q1 2026, accelerating from 39% the prior quarter. The number looks strong on the surface. The caveat is that capacity is being diverted to power Microsoft Copilot products, which means standard Azure workloads are competing internally for the same compute. Microsoft is also navigating the OpenAI exclusivity change. The deal still makes Microsoft OpenAI's primary cloud partner through 2032, but the exclusive lock on model distribution ended. Both companies framed it as a win. Structurally, it is a meaningful shift in leverage.
GCP (Google Cloud) Google Cloud crossed $20 billion in quarterly revenue for the first time in history, growing 63% year over year against a $12.3 billion comparison period. Alphabet beat Wall Street estimates broadly, and Cloud was the headline driver. The caveat Google flagged on its earnings call is worth paying attention to: growth was capacity-constrained. Demand is outrunning supply. Google is building fast, and the $710 billion CapEx commitment across the four major hyperscalers is the answer to that problem. CapEx for Big Tech is now projected to exceed $1 trillion annually by 2027.
AI Model Roundup
OpenAI Two major OpenAI stories hit this week and they pull in opposite directions. On the product side, the AWS Bedrock partnership makes OpenAI models more accessible to more developers than at any point in the company's history. On the governance side, the Musk trial opened in Oakland. Musk is seeking $134 billion, arguing that OpenAI's conversion from nonprofit to for-profit entity broke the original founding commitments. OpenAI listed the lawsuit as a business risk in its IPO materials. The trial outcome will not change the technology, but it will shape how AI governance clauses are written in enterprise contracts for the next decade.
Anthropic Anthropic was not in the Pentagon deals. That choice reflects their stated position on AI safety, but it comes with real business consequences. The AI resource war is real. Compute is constrained, training runs are expensive, and the companies getting classified government contracts are also getting preferred access to infrastructure. Anthropic is running a different strategy, and it is worth watching whether that strategy holds as the capital requirements grow. The gap between Anthropic's compute access and OpenAI's is widening, not narrowing.
Google AI Google's AI week was strong across the board. Gemini models are shipping across Google products, from Maps to design tools. Google is part of the Pentagon classified network agreements. And Google Cloud's 63% growth reflects real enterprise demand for Gemini-based products, not just infrastructure spend. The question worth asking is whether Google's AI advantage in search and consumer products translates into AI platform wins in enterprise. The Q1 numbers suggest it is starting to.
The Pattern I'm Watching
I have watched infrastructure consolidation happen three times at scale in the last thirty years. It happened with enterprise software in the late 1990s, when the ERP market went from dozens of vendors to SAP and Oracle. It happened with public cloud in the 2010s, when three providers absorbed most of the market. And it is happening again right now with AI infrastructure.
The Pentagon's vendor list this week is not just a government procurement story. It is a signal that the consolidation phase for AI infrastructure has started. When a government agency is willing to put classified national security workloads on a vendor's platform, that vendor has cleared a trust threshold that takes years to rebuild for a challenger. Microsoft, AWS, and Nvidia are not just winning contracts. They are building institutional moats that will be as durable as the ones Oracle and SAP built in the 1990s, maybe more so because the switching costs in AI workloads are higher than they were in ERP. You cannot easily retrain a frontier model, swap your CUDA-optimized inference stack, or rebuild your classified IL-6 deployment for a new vendor without years of work.
The wild card is governance. The Musk trial, the Anthropic principled refusal, the CVE-2026-31431 patch window exposure all point to the same question: as AI gets embedded deeper into critical infrastructure, who is actually accountable when something goes wrong? In the ERP era, accountability was contractual. In the cloud era, it became shared responsibility. In the AI era, nobody has figured out the governance model yet. The companies that do will have a second moat on top of the infrastructure moat. That is the pattern worth watching over the next ten years.
Here is the question I keep coming back to: if you are an enterprise CTO today, and you are watching the Pentagon consolidate around seven vendors, does that make you more confident in your own AI vendor selection, or does it make you more nervous about being locked in to a list someone in Washington approved?
Weekly AI and cloud breakdowns from someone who's been in the game since the early days of the internet. No ads. No filler. The signal.