AI in the News

AI in the news: week of May 3, 2026

Big Tech Q1 prints AI as the lead growth engine, Microsoft Agent 365 hits GA as a paid control plane, GPT-5.5 carries momentum into enterprise rollouts, EU AI Act trilogue stumbles, and Challenger calls AI 26% of April's cuts. Capital is flowing into compute, not labor.

Sid Smith

03 May 2026 • 7 min read

What this week actually changed: Big Tech's Q1 prints made it official (capex up, headcount down, AI cited on both) and Microsoft turned the next product category (the agent control plane) into a $15/user/month line item.

The week ending Sunday May 3 was a capital-and-control-plane week. Big Tech printed Q1, and the prints were AI prints. Microsoft turned Agent 365 on as a paid product. OpenAI's GPT-5.5 momentum from April 23 kept rolling into enterprise rollouts. Brussels failed to land the AI Act trilogue with the August deadline in plain sight. And Challenger Gray's April layoff report dropped a number that's hard to look away from.

Big Tech Q1: capital is flowing into compute, not labor

Alphabet, Microsoft, Meta, and Amazon all reported the week of April 27. The throughline: cloud-AI is the growth engine, and the capex guides went up, not down. Alphabet raised 2026 capex to $180–190B with Google Cloud +63% YoY to $20B in the quarter. Meta revised 2026 capex to $125–145B on $47B in revenue (+33%). Amazon posted AWS at its fastest growth rate in 15 quarters. Azure stayed in the +40% band.

The numbers matter less than the posture. Every hyperscaler is now guiding to higher capex than they were six months ago, in the same quarter they're announcing layoffs. That combination (record capex, headcount reductions) is the shape of the 2026 AI economy in one sentence. Capital is flowing into compute, not into labor.

What I'll watch: whether Q2 prints hold the line. Current consensus says the AI capex cycle is self-sustaining because demand keeps showing up. The contrarian read is we're 18 months into a hyperscaler arms race where the marginal dollar of compute spend is yielding less marginal revenue than the dollar before it. Neither view is provable from a single quarter. Two more prints and the trajectory becomes legible.

Microsoft Agent 365 GA, the agent control plane just became a product

May 1. Microsoft made Agent 365 generally available at $15 per user per month, with an E7 bundle at $99/user/month that wraps Copilot + Agent 365 + Entra Suite + E5 security. The pitch is a three-pillar control plane: observe, govern, secure. Includes registry sync with AWS Bedrock and Google Cloud so the inventory is cross-cloud, not just Microsoft-internal.

This is the most important enterprise-AI product launch of the year so far and the press is mostly missing why. Agent 365 isn't competing on agent capability, it's competing on agent control. The bet is that the enterprise problem of 2026 isn't "can we build agents" but "we have 400 agents running and nobody knows what any of them are doing." Microsoft is selling the control plane that sits above every agent stack, including its competitors'. That's a different positioning than "use Copilot for everything" and a smarter one.

Where I push back: the control plane should be open. Agent 365's registry sync with Bedrock and Google Cloud is the right shape, but the governance work I keep arguing for doesn't depend on a Microsoft product to exist. An open inventory-and-policy spec for enterprise agents would let every cloud, every framework, and every on-prem deployment participate without paying $15 per user per month. Microsoft has the distribution to lock this in as proprietary. Whether the open alternatives form fast enough is the question.

GPT-5.5 in the wild, the model is good

OpenAI shipped GPT-5.5 on April 23, so it lands inside this window for enterprise availability, the GPT-5.5 Pro parallel-reasoning variant, and Codex on the combined stack. The framing in the press was "superapp", agentic coding plus computer use plus knowledge work in one model. Big jumps on agentic-coding benchmarks. Real progress on long-horizon task completion.

The model is good. I've been running 5.5 against 5 for the last week on the same agent harnesses and the completion-rate delta on multi-step tool-use tasks is the largest single-version jump I've seen from OpenAI in a year. The architectural and training-stack details are still mostly opaque, but the behavioral evidence is clear: longer agentic loops are more reliable.

The thing worth flagging is the chip story attached to it. OpenAI is reportedly in talks with MediaTek and Qualcomm on a custom chip for an OpenAI-branded device. Take that with appropriate salt (the device is reportedly still 12+ months out) but the strategic direction is consistent: OpenAI wants the full stack, model through hardware. The hosted-frontier-AI bet keeps getting more vertical.

EU AI Act trilogue stumbled, then landed

April 28. The second political trilogue between the European Parliament, Council, and Commission ended without agreement on the AI Act simplification package. The high-risk-AI provisions enter force August 2. The simplification proposal was supposed to land before then.

A breakthrough did come on May 7 (the next week) that postponed national regulatory sandboxes to August 2027, tightened transparency for AI-generated content to a 3-month grace period, and added a new prohibition on non-consensual sexual and intimate AI content. So the news cycle for this story actually crossed the week boundary.

Two reads. First, process-wise, this is fine. Legislative trilogues run hot in the final weeks and one failed session is not a crisis. The May 7 agreement got most of what mattered. Second, structurally: the EU is now in the position of debating whether to delay its own landmark AI law while the US runs a state-by-state patchwork and China runs the opposite playbook. The convergence-or-divergence question I flagged last fall keeps trending divergence. The EU AI Act and SB 53 don't share enough taxonomy to cross-comply against, and the compliance overhead for multinational labs keeps growing.

The non-consensual intimate content prohibition is the right move and overdue. It's also a tacit acknowledgment that the existing AI Act didn't cover the most obvious harm vector cleanly enough. Worth watching whether other jurisdictions copy it.

Labor: Challenger says AI was 26% of April's cuts

May 1. The Challenger Gray report dropped. AI was cited as a driver in 26% of April's announced layoffs, the highest share since they started tracking the category. Q1 2026 tech-sector layoffs hit roughly 80,000, with about half AI-cited per Tom's Hardware's read. Meta announced 10% / ~8,000 cuts and Microsoft another tranche, pushing the combined two-company number past 20,000.

Sam Altman's framing at the India AI Impact Summit was that there's "some AI washing where people are blaming AI for layoffs they would otherwise do, and then some real displacement by AI." That's a more honest framing than I expected from the OpenAI CEO and I'll credit it. He's right that both are happening. I think the second category (real displacement) is the larger one, and growing.

The displacement is real and it's accelerating faster than I expected. The pace is what's wrong. Companies aren't cutting because the AI is ready for the work; they're cutting because the AI narrative is convenient and the markets reward the cuts. The Meta and Microsoft announcements landed in the same week as raised capex guides, the headline reads "spend more on compute, employ fewer humans." The financial logic is plain and the short-term incentive is loud.

The sustainable model is human+AI collaboration. The companies that figure out the collaboration outperform the companies that just cut, but the headcount still shrinks under collaboration, it just shrinks less and shrinks well. I'd rather be wrong about how fast this moves than be the person caught flat-footed. More on this here.

Smaller items worth tracking

Google-Anthropic deal closes the loop. Google's up-to-$40B investment in Anthropic, announced April 24-25, hangs over this week. $10B cash up front: $30B contingent on milestones: $380B valuation. Google now has meaningful equity exposure to both its main competitor (via OpenAI's compute history) and its main alternative (Anthropic). The hyperscaler-lab capital structure keeps getting more entangled.
DeepSeek V4 pricing. DeepSeek reportedly cut V4-Pro pricing as low as $0.0036 per million tokens. Treat the headline number with skepticism on apples-to-apples comparison, but the pricing trajectory from the Chinese labs is real and downward.
Anthropic Project Glasswing. Anthropic announced a preview that identified thousands of zero-days across major OSes and browsers in weeks of testing. Worth tracking. The dual-use surface is enormous and the disclosure-vs-weaponization question matters more here than in most AI launches.

What to watch next week

Three threads. The capital story and the labor story are the same story. Hyperscalers raised capex guides and announced layoffs in the same earnings cycle. That's not a coincidence and it's not noise. Capital is flowing into compute, not labor, and the executive class is being plain about it. The discomfort that creates is the right discomfort.

The control plane is the next product category. Agent 365 going GA at $15/user/month is the proof point. Once enterprises have hundreds of agents, the problem isn't agent capability, it's agent governance. Microsoft is first to ship a paid control plane. The open alternatives have a narrow window to form before this gets locked in as proprietary.

And the regulatory frame keeps diverging across jurisdictions. EU debating delays, US running a state patchwork, China running its own playbook, and multinational labs needing to comply with all of them. SB 53's first reporting cycle is approaching, the EU AI Act's August 2 deadline is approaching, and the cross-jurisdiction reconciliation is not happening.

Next Sunday: the May 7 EU AI Act agreement details, the Agent 365 first-week adoption signal, and whatever earnings stragglers land.