Good morning,
We cover hyperscalers' Q1 earnings, the rewriting of the Microsoft–OpenAI partnership, as well as new JV projects on both the OpenAI and Anthropic sides. Plus the usual top news roundup: another round of tech layoffs, DeepSeek V4 release, and Germany’s Aleph Alpha and Canada’s Cohere merger.
Hyperscaler Q1: capex guides raised again
2025 Q4 calls from big tech raised an increasing focus on CAPEX with massive YoY increases, so the question coming into earnings week was whether 2026 guides would hold or move further up. And… they moved up.
Once you stack the 29 April prints from Microsoft, Alphabet, Meta and Amazon, combined 2026 AI infrastructure spend now sits at $700–725bn, materially above the initial January guides.
2025 Actual CAPEX | Guidance CAPEX Jan ’26 | Guidance CAPEX Apr ’26 | |
|---|---|---|---|
Amazon | $134.7 billion | $200 billion | $200 billion |
$91.5 billion | $175-185 billion | $180-190 billion | |
Meta | $72.2 billion | $115-135 billion | $125-145 billion |
Microsoft | $118 billion | $140+ billion (estimate) | $190 billion |
TOTAL | $416.4 billion | $640+ billion | $695-725 billion |
Component-price inflation is the new driver. Microsoft attributed roughly $25bn of its $120bn 2026 capex to component-price inflation alone; Meta cited the same factor in lifting its guide to $125–145bn. The "compute will get cheaper, margins will improve" thesis that underpinned much of the bull case is, for the moment, running in reverse. This aligns with what we covered two weeks ago on Data Center slowdown.
The "AI profit loop". As Fortune pointed out, close to half of Alphabet's record $62.6bn quarter (~$28.7bn) came not from operations but from marking up the value of its private stake in Anthropic. Amazon disclosed $16.8bn in pre-tax gains on its $8bn Anthropic investment. Microsoft has recognized $5.9bn of OpenAI-related gains over the past nine months. Critics call this a circular paper-accounting loop: hyperscalers fund the labs, the labs spend the money back on hyperscaler compute, and the hyperscalers revalue their lab stakes upwards based on funding rounds in which they themselves are the largest cheque-writers. Bulls counter that the underlying revenue is real — Anthropic's run-rate is roughly $30bn officially and reportedly closer to $40bn by mid-year. With Anthropic and OpenAI in talk to IPO, these mark-ups will only grow.
Another datapoint worth noting: hyperscaler debt issuance is now forecast at ~$175bn in 2026, around six times the five-year average. Alphabet sold a 100-year "century bond" (the first by a tech company since Motorola in 1997). The financing of the build-out is becoming as interesting as the build-out itself; and perhaps riskier.
Why it matters: Two implications shift our read of Big Tech as a leading indicator. First, the "AI profit loop" complicates earnings-quality assessment: a meaningful share of the printed profits is paper marks on private stakes whose value depends on the next funding round. Second, the build-out is increasingly debt-funded at a scale most observers haven't internalized. If AI revenue falls short of projections, financial stability will not be contained to the technology sector alone, and we risk seeing the AI bubble pop.
OpenAI and Anthropic the fight continues
While earnings dominated the headlines, another strategically consequential event unfolded in late april as Microsoft and OpenAI announced a restructuring of their partnership.
Four things changed materially. Microsoft's IP licence became non-exclusive through 2032 — GPT-5.4 hit AWS Bedrock the day after the announcement. The AGI clause — the contractual provision that would have unwound parts of the partnership upon Microsoft declaring OpenAI had built AGI — is gone. Revenue-share payments from OpenAI to Microsoft are capped but extended to 2030. And Microsoft retains its ~27% stake (~$135bn) but now functions, in practice, as a shareholder rather than an exclusive distribution partner.
In parallel, Anthropic is reportedly in talks to raise a $40–50bn round at a ~$900bn valuation and is launching $1.5bn enterprise JV with Blackstone, Hellman & Friedman and Goldman Sachs. The structured vehicle will be backed by the three plus Apollo, GA, GIC, Leonard Green and Sequoia, to deploy Claude inside large enterprises using Palantir-style forward-deployed engineers. The same day, Bloomberg reported that OpenAI is working on a parallel $10bn vehicle ("The Development Company") backed by TPG, Brookfield, Advent and Bain. Same model: bringing senior engineers in-house to clients to operationalise frontier AI. Worth noting: PE, not VC, are taking the lead on AI deployment, and the largest financial-services GPs are visibly positioning around it.
As the public perception and money fight continues, Counterpoint Research quietly published a Q1 snapshot that shows Anthropic has now overtaken OpenAI in global LLM revenue (31.4% vs. 29% in Q1) on roughly one-seventh the user base — implying ARPU around $16.20 vs. OpenAI's $2.20.
Why it matters: The idea that "OpenAI is #1, Anthropic is the safety shop" might be coming to an end. Anthropic now leads on revenue, on price-per-user economics, and increasingly on enterprise mindshare in regulated industries — although OpenAI seem to still lead in B2C. The Microsoft-OpenAI rewrite also signals that the era of single-cloud AI lock-in is over: cloud strategy and AI-vendor strategy are decoupling. For procurement and IT roadmap planning, that means we should expect more multi-cloud, more multi-vendor architectures, so messier governance: with five frontier labs serving every major cloud, accountability for model behaviour is becoming distributed.
More Top News
Another quarter, another tech layoffs, hitting ~80,000, with 48% AI-attributed. Amazon (16,000), Oracle (30,000), Meta (8,000), Salesforce (4,000), Microsoft (~7% buyouts), Intel (15%, plus cancelled Germany and Poland fabs), all in the window. Sam Altman publicly conceded some of these are "AI-washed" — i.e. unrelated cuts being attributed to AI for narrative convenience. Counter-view: Earlier this year, IBM had announced plans to triple entry-level hiring in 2026, betting against the consensus that juniors are the first to go. Let’s wait and see.
DeepSeek V4 (finally) launches with native Huawei-chip support. On 24 April, DeepSeek released V4 Flash and V4 Pro open source, with a 1-million-token context, a "Hybrid Attention Architecture" and native support for Huawei Ascend silicon rather than CUDA. It is the first frontier-class system designed from the outset to run without US-controlled accelerators. Stanford's AI Index now puts the US-China model gap at ~2.7%; the hardware decoupling, not the model quality, is the bigger story. Following on our previous coverage of the Z.ai/Huawei thread, China's two-track strategy is clearly working.
xAI x Cursor. Musk’s rocket company SpaceX – which merged with xAI earlier this year – has concluded an agreement with Cursor giving it the right to acquire the AI coding startup for $60 billion later this year, or pay $10 billion for the companies’ work together. The costly tie-up is aimed at developing “the world’s best coding and knowledge work AI,” SpaceX said.
Cohere and Aleph Alpha merge to form a sovereign-AI champion. Announced 25 April, the Canadian-German tie-up is backed by both governments at a ~$20bn combined valuation, with a Schwarz Group-led Series E. The pitch: defence, energy, healthcare and public-sector customers wanting a non-US "sovereign" stack. The "sovereign AI" category is becoming concrete.
Mythos cyber-capability evaluations land. Following on our previous edition, AISI published its full Mythos Preview evaluation — Mythos succeeded on 73% of expert-level CTFs. The Bloomsbury Intelligence and Security Institute concluded that Mythos is more likely, in the near term, to raise cyber risk via faster vulnerability discovery than the patch cycle can absorb. The WSJ reported that the White House opposes any wider Mythos release on national-security grounds.
Vercel breached via Context.AI supply-chain attack. Disclosed in late April: compromised Context.AI OAuth tokens were used to access Vercel internal data, then offered for sale on BreachForums by ShinyHunters for $2m. Vimeo was caught up via a parallel Anodot/Snowflake compromise on the same day. A reminder that the AI tooling layer is now itself a meaningful attack surface and one most enterprises have not yet inventoried.
Oups. An autonomous Cursor coding agent running Claude Opus 4.6 deleted PocketOS's entire production database in nine seconds while "fixing" a credential mismatch.
Meta to track employee screens and keystrokes for agent training. Reuters reported that Meta Superintelligence Labs will instrument employee workstations to gather training data for its agent products. A precedent that will recur in labour and data-protection cases and a question every employer with an internal AI-agent ambition will eventually face.
OpenAI rolls out advertising in ChatGPT. The ads in ChatGPT pilot proved very successful crossing $100m ARR in under six weeks; no wonder then that OpenAI is now expending its reach. A material shift in OpenAI's revenue model and the start, perhaps, of AI assistants becoming a media business. Like we expressed previously, this shows that OpenAI main MOAT is its distribution.
Tool to try
With an increasing focus on B2B and turning models into a profit machine, we are sadly talking less and less about creative models. This doesn’t mean they are not getting better! OpenAI released Images 2.0 at the end of April, and it is phenomenal. Not only is the new model amazing at generating realistic images, but it finally fixes major issues with text generation. I definitely recommend giving it a try!
As a famous AI reviewer would say: What a time to be alive!