This week felt like a turning point. Eye-watering capital spending, regulators stepping up, models that companies are afraid to ship without guardrails. Let's get into it.
COVER STORY
The $725 Billion Dollar Bet That Everyone Is Forced to Make
If you wanted a single chart to explain the mood in tech right now, look at the capital expenditure numbers from Meta, Amazon, Microsoft, and Alphabet. Together they have signaled roughly $725 billion in capex for 2026 — a jump of more than 75% year over year, almost all of it pouring into data centers, custom silicon, GPUs, and frontier models. It is the largest infrastructure buildout in the history of software.
What makes the moment feel strange is the other side of the same balance sheet. Meta is moving forward with around 8,000 layoffs this month. Amazon has cut roughly 30,000 roles in recent months. Microsoft is offering voluntary buyouts to about 125,000 employees. The companies pouring the most money into AI are also the ones reshaping their workforce around it — the productivity story is no longer theoretical.
The real question this week is not whether AI is worth the spend. The question is whether anyone can afford not to spend. When your competitors are committing hundreds of billions to compute, sitting on the sidelines starts to look like a strategic error rather than a prudent one.
What to watch: If revenue from agentic AI deployments does not start materially closing the gap with capex by H2 2026, expect the first real conversations about a slowdown.
POLICY & REGULATION
Europe Blinks, But Only a Little: Inside the AI Act Omnibus
On May 7th, the European Parliament and Council reached a political agreement on amendments to the AI Act, packaged inside what Brussels is calling the AI Act Omnibus. Compliance deadlines are being stretched, certain industrial AI use cases are being carved out, and generative AI watermarking rules are being phased in more gradually.
The Omnibus deal does not abandon the framework. It softens the timing and trims the friction for industry, while adding new explicit prohibitions on so-called nudifier applications that generate non-consensual intimate imagery.
Governance signal: The Act still becomes fully applicable on August 2nd. Anyone building or buying AI for the European market should treat the next six weeks as the window to finalize their compliance posture.
THE DEBATE
Should Governments Test AI Models Before You Do?
This week Microsoft, Google, and xAI confirmed they will let the US government test their frontier AI models before public release. OpenAI agreed to give the EU early access to a cybersecurity variant of GPT-5.5. Anthropic, by contrast, is still holding out on sharing its newest model called Mythos, which it considers to carry serious cybersecurity misuse risk.
The harder question is what counts as a meaningful evaluation. Frontier models are so capable that testing them in a sealed environment for a few weeks tells you very little about how they will behave once millions of people, agents, and downstream apps start composing them in unexpected ways.
Counterpoint: Anthropic's restraint with Mythos may turn out to be the more durable signal. A lab voluntarily slowing distribution of a model it considers too dangerous is the kind of behavior the next era of AI governance will need to reward, not penalize.
QUICK TAKES
- Novo Nordisk and OpenAI go all-in on AI for drug discovery. A sweeping partnership to embed OpenAI across drug discovery, clinical trials, manufacturing, and supply chains — full deployment targeted by year end.
- Cerebras pops 68% on its Nasdaq debut. The wafer-scale chipmaker hit a $95B valuation on day one — the clearest public market signal yet that investors will pay a premium for serious Nvidia alternatives in AI compute.
- Anthropic taps SpaceX Colossus 1 for compute. Confirmed May 6th — all available compute from the Memphis Colossus 1 data center, alongside expanded partnerships with Google and Broadcom.
- Google races to put Gemini at the heart of Android. New features complete multistep tasks across apps, browse the web, fill forms, and generate custom widgets through natural language — widely read as preparation for Apple's upcoming AI reboot.
If one theme tied it all together this week, it is that the AI conversation has finally moved past whether the technology is real. It is now about who pays for it, who governs it, and who gets to decide which models the world is allowed to use.
— Rudy Shoushany, Editor · SIGNAL Issue #6 · May 17, 2026
