AI Week in Review: $700B Capex, GPT-5.5, and a Stalled EU AI Act (April 24 – May 1, 2026)

Weekly AI digest — April 24 to May 1, 2026.

If a single week could capture the strange duality of the AI cycle in 2026, this was it. On the same days that Big Tech reported eye-watering capital expenditure plans north of $700 billion for the year, Brussels could not agree on whether to delay the AI Act, China’s DeepSeek shipped a new flagship model, and OpenAI pushed ChatGPT a step closer to becoming a single agentic “super app.” The question is no longer whether generative AI will industrialise — that battle is over. The question is who will pay for it, who will regulate it, and whether the productivity payoff will arrive before investor patience runs out.

1. The $700 billion bet: hyperscaler earnings reset the spending floor

Wednesday April 29 was the most consequential day of the AI financial calendar. Within a few hours of the closing bell, Alphabet, Microsoft, Meta and Amazon all reported quarterly results — and all four lifted their AI infrastructure forecasts. Combined 2026 capex is now tracking between $650 and $700 billion, the largest concentrated infrastructure cycle in the history of the technology industry.

The individual numbers tell the story:

• Alphabet raised full-year capex guidance to a range of $180–190 billion and warned that 2027 spending will “significantly increase.” Google Cloud revenue grew 63% year-over-year to $20 billion, and revenue from products built on Alphabet’s generative AI models was up almost 800%.
• Microsoft guided fourth-quarter capex above $40 billion and now expects roughly $190 billion for the full year.
• Amazon confirmed an annual capex envelope of $200 billion, the largest of the four. AWS revenue reached $37.6 billion, growing 28% — its fastest pace in fifteen quarters.
• Meta raised its capex range to $125–145 billion, up from $115–135 billion. Revenue grew 33%, the company’s fastest expansion since 2021, but the stock fell roughly 7% after-hours as investors balked at the spending lift.

The market verdict was strikingly uneven: Alphabet rose almost 7% after the print, while Meta sold off and Microsoft was essentially flat. The single dividing line was clarity of return on investment. Where executives could point to AI revenue already on the books — Google Cloud being the cleanest case — the spending was rewarded. Where the case rested on future agents, future enterprise adoption or future ad-targeting gains, investors flinched.

The combined capex of Alphabet, Microsoft, Amazon and Meta in 2026 will exceed the entire annual GDP of countries like Switzerland or Saudi Arabia. We are no longer talking about a tech cycle. We are talking about an industrial buildout on the scale of national infrastructure programs.

2. The model race: GPT-5.5, DeepSeek V4, Qwen 3.6

The capex numbers are the macro story. The micro story is that frontier models keep getting better — and cheaper — every few weeks.

OpenAI released GPT-5.5, a mid-generation upgrade explicitly positioned as the foundation of a unified ChatGPT experience: a single product that combines conversation, coding, browsing and a native computer-use mode. Alongside the model, OpenAI launched workspace agents in ChatGPT Business, Enterprise and Education tiers, allowing teams to build, share and run agents that act across Slack, Gmail and other workplace tools. The strategic message is unambiguous: ChatGPT is no longer a chatbot, it is an operating layer for knowledge work.

OpenAI’s commercial trajectory continues to defy expectations. The company has now passed $25 billion in annualized revenue and is reportedly preparing the groundwork for a public listing as early as late 2026.

The week also confirmed that the frontier is genuinely contested. China’s DeepSeek rolled out preview versions of its V4 Flash and V4 Pro series, with top-tier coding-benchmark scores and notable gains on agentic and reasoning tasks. Alibaba‘s Qwen team released a preview of Qwen 3.6 Max, billed as the lab’s most capable model yet, with similar improvements in coding, reasoning and agentic execution. Together with Anthropic’s rumored “Claude Mythos” frontier model and Google’s recently shipped Gemini 3.1 Flash-Lite — which delivers 2.5× faster responses at $0.25 per million input tokens — April 2026 has now become the densest model-release window in the industry’s history.

What unifies these releases is not raw IQ — gains on classical benchmarks are increasingly marginal — but agentic readiness: longer context, native tool use, robust computer control, and pricing aggressive enough to make 24/7 deployment economically viable.

3. Money flows: Google’s $40B Anthropic deal and the Roze IPO

The AI capital cycle is not just a hyperscaler story. Two announcements this week reshape the financing map.

First, Google committed up to $40 billion in cash and compute to Anthropic. Combined with a fresh $5 billion top-up from Amazon, and a broader compute agreement under which Anthropic is expected to spend up to $100 billion for around 5 gigawatts of capacity over time, the company is now structurally tethered to two of the three largest cloud providers in the world. Anthropic’s annualized revenue is approaching $19 billion, less than OpenAI’s but growing on a similar curve.

Second, SoftBank is reportedly preparing to spin out and list a new AI-and-robotics company called Roze, targeting a valuation of up to $100 billion. Roze’s stated focus is the physical buildout of AI infrastructure — using robotics to help construct the data centers the rest of the industry is now committed to building. It is a telling sign of where the cycle has moved: from training the models, to powering them, to physically building the buildings that will house them.

4. Regulation: Brussels stalls, Washington fragments

While capital was flowing, the regulatory side of the ledger was conspicuously stuck. On the night of April 28 to 29, after twelve hours of negotiations, the European Parliament and Council failed to reach a common position on the proposed Digital Omnibus reforms to the AI Act. Talks are scheduled to resume in May.

The central issue is timing. Member states and lawmakers had previously aligned on postponing the high-risk AI compliance deadlines — Annex III systems would slip to December 2, 2027, and Annex I (AI embedded in regulated products) to August 2, 2028. But the political coalition behind that delay is fraying, with civil-society groups arguing the postponement amounts to deregulation by stealth and several capitals demanding firmer guarantees on transparency and incident reporting in exchange for any delay. The August 2026 compliance deadline for general-purpose models with systemic risk is unaffected and looms within months.

On April 21, separately, the European Commission released €63.2 million to support AI innovation in health and online safety — a small amount in the context of $700 billion of corporate capex, but politically significant as a signal that Brussels still intends to fund European players rather than rely solely on rule-making.

In the United States, the regulatory picture remains state-by-state. New York amended its frontier AI framework in March to align more closely with California’s transparency-and-reporting model under the RAISE Act. There is still no federal AI statute, and the gap between American capital intensity and American regulatory intensity continues to widen.

5. The product layer: agents go mainstream

Beyond the headline labs, the week brought a flurry of product news that quietly confirmed the same direction of travel.

Meta disclosed that its business AI assistant now facilitates 10 million conversations per week across WhatsApp, Messenger and Instagram — a credible adoption number for an enterprise product launched less than a year ago, and a signal that small and mid-sized merchants are integrating AI customer service faster than enterprises are integrating internal copilots.

Google’s cloud unit, at its annual Las Vegas conference, unveiled a set of tools for building, deploying and tracking AI agents inside companies, including a dedicated inbox where bots post status reports to their human colleagues. The framing — agents as digital coworkers with their own message inbox — is increasingly the consensus interface metaphor across vendors.

Underneath all of this sits a subtle but important shift: the industry is converging on the language of “software as employee” rather than “software as tool.” Whether that framing survives contact with HR law, audit requirements and basic enterprise risk management is one of the open questions of 2026 and 2027.

What to watch next week

Three things to follow into the first week of May:

1. Apple’s earnings, scheduled for May 1, will be the next read on whether the only major US tech company without a flagship frontier model can keep monetising AI through device-level integration.
2. The next round of EU Digital Omnibus talks, expected mid-May, will determine whether the AI Act delay holds — and therefore the actual cost of European compliance for every model provider.
3. The US semiconductor supply chain, which is increasingly the binding constraint behind the $700 billion capex number. Any disruption — export controls, packaging capacity, HBM availability — will hit hyperscaler guidance before it hits headlines.

Bottom line

This week did not produce a single, dramatic AI breakthrough. It produced something arguably more important: confirmation that the industrial phase of the AI cycle is now fully underway. Capital is committed at unprecedented scale. Models are getting better and cheaper on a roughly monthly cadence. Agents are crossing from demo to deployment. And the regulatory and political response — in Brussels, Washington and Beijing — is visibly lagging the technical and financial reality on the ground.

The strategic implication for anyone building, investing or governing in this space is straightforward: the window in which AI was a debate about whether is closed. The debate that matters now is about distribution — of compute, of returns, of regulatory burden, and of the productivity gains that the $700 billion bet implicitly promises.

Sources: Q1 2026 earnings releases (Alphabet, Microsoft, Amazon, Meta), TechCrunch, Bloomberg, Fortune, CNBC, Reuters, IAPP, Holland & Knight, European Commission press materials.