TechLook
AI infrastructure just became a national budget line. That's the story of today.
While the conversation about AI often stays停留在models and benchmarks, the real action this week is at the foundation layer — who is paying for it, who is controlling access, and what happens to the humans in the middle of an acceleration that won't slow down for anyone.
Three things hit at once: hyperscalers are on pace to spend more in 2026 than every non-tech company in the S&P 500 combined. The White House wants pre-release access to frontier AI models. And the number of announced tech layoffs hit 81,000 in a single quarter — all while AI budgets hit record highs. These aren't contradictory. They're the same story.
AI Capex Hits $805B — More Than the Rest of S&P 500 Combined
The number that should stop every founder and builder mid-scroll: hyperscalers Amazon, Alphabet, Meta, Microsoft, and Oracle are projected to spend $805 billion on AI infrastructure in 2026 alone. Morgan Stanley raised the forecast, and the 2027 number is already penciled in at $1.1 trillion.
To put that in context: the 2026 figure could match the combined capital expenditure of every non-technology company in the S&P 500 for all of 2025. This isn't building software. This is building civilization-scale infrastructure.
Here's everything you need to know:
- Morgan Stanley raised its 2026 capex forecast for the five hyperscalers after updated guidance from all major players
- 2027 forecast sits at ~$1.1 trillion — a figure that would have seemed fantastical two years ago
- AI infrastructure has become a nation-state arms race, not a Silicon Valley project
- Data centers, power infrastructure, and custom silicon are the bottleneck — not model capability
- This spending is going to compress margins across the industry even as it enables new product categories
The implication for founders is direct: the underlying compute layer is being locked up by companies with trillion-dollar balance sheets. If your product depends on inference at scale, the unit economics are going to be decided by these players' pricing decisions, not by your architecture choices. Build accordingly.
OpenAI Launches $10B Deployment Company — Private Capital Now Runs AI Rollout
While OpenAI's core model work stays in-house, the company's new joint venture — called "The Deployment Company" — is something else entirely. It's a $10 billion structure with private equity firms Bain Capital, Brookfield, and TPG, designed to deploy OpenAI tools across their massive portfolio companies.
Here's everything you need to know:
- The vehicle is designed to inject OpenAI into PE portfolio companies at scale — faster than enterprise sales could move
- Blackstone, Goldman Sachs, and Sequoia are running a parallel structure for Anthropic's mid-market push
- These deals effectively give private capital a direct channel into AI adoption without going through traditional enterprise software sales
- Anthropic's new services company pairs Anthropic engineers with Blackstone's team to deploy Claude into mid-sized businesses with limited technical staff
- Coverage areas include documentation, coding, prior authorizations, and compliance reviews — real operational work
The open question: is this actual operating help or expensive consulting with better branding? For mid-market companies with skeleton crews, having Anthropic engineers embedded might be the difference between a real deployment and a stalled pilot. For larger enterprises, it may just be a new layer of vendor markup.
If these structures work — and generate the revenue these firms are projecting — expect every major AI lab to get a services subsidiary. The model-as-a-service era is already giving way to model-plus-implementation-as-a-service.
White House Wants First Look at Frontier AI Before Public Release
The Trump administration is considering a working group that would review powerful AI models before they go public — specifically targeting models with potential national security, cyber, or defense applications. Officials met last week with Anthropic, Google, and OpenAI following Anthropic's disclosure that its Mythos model was too capable at finding software vulnerabilities to release publicly.
Here's everything you need to know:
- The ask: government first access to models that could affect critical infrastructure, cybersecurity, or weapons development
- Anthropic has already been blacklisted from the Pentagon's most sensitive networks for refusing a broad "any lawful use" clause
- The company is fighting the Pentagon over a $200M contract dispute and reportedly suing over the contract terms
- Critics argue Anthropic's "too dangerous to release" framing may be overstated — UK AISI testing found OpenAI's GPT-5.5 actually scored higher than Mythos Preview on cybersecurity tasks (71.4% vs 68.6%)
- Both models became the first to breach AISI's simulated enterprise network test
- The risk: "review" scope tends to expand once Washington gets comfortable with access
What this means for founders building at the frontier: the regulatory line is being drawn in real time, and it's not stable. If you're training on code generation, vulnerability research, or anything adjacent to cyber-offense — expect more scrutiny, more disclosure requests, and potentially new licensing regimes. The era of shipping first and asking questions later is closing.
81,000 Tech Layoffs in Q1 — The AI Boom Is a Labor Compression Story
The jobs numbers don't match the narrative. While AI budgets hit all-time highs, layoffs.fyi shows 81,747 tech layoffs announced in Q1 2026 — up 580% in months. The AI boom is not creating jobs in the same places it's destroying them.
Here's everything you need to know:
- Q1 2026 layoffs running at 81,747 and accelerating
- 2026 AI capex and automation budgets are at record levels
- The roles being cut are not being replaced by AI-related hiring in the same companies
- The disconnect is most acute at mid-level: managers, coordinators, and operational roles that existed to bridge human workflow are being compressed out
- New hiring is concentrated in research, infrastructure, and sales — not the broad middle of corporate org charts
This is the uncomfortable stat that gets left out of every AI progress report. The technology is advancing faster than any previous wave. The displacement is also moving faster. Founders building AI-native products are part of this — either as the cause or as the people who need to think harder about what they owe the humans in their ecosystem.
One Person. Multiple Agents. $12K in Six Days.
While the enterprise deals make headlines, the most striking signal in today's news is smaller: a founder who quit their job, built an AI-powered SaaS alone, and reached $12,000 in monthly recurring revenue within six days of launch using two devices running AI agents — one building features, the other reviewing PRs, running tests, checking security, and blocking bad code.
Here's everything you need to know:
- Solo founder used two devices running AI agents in parallel — one handled development, the other quality control and security review
- This is the "couch-run agent" model: founder operates from home, delegates execution entirely to agent teams
- Agent infrastructure tools like OpenClaw 5.3 now have file transfer for paired nodes and improved live control (/steer, /side commands)
- The AI Marketing Engineer role is emerging as a distinct discipline: prototype in agentic tools, run real tasks, let corrections become reusable skills, then move stable flows into scripting and cron
The story from India's AI Festival in Goa reinforces it: PB Teja, an OpenClaw core developer, built Mission Control HQ on his phone using an agent team for marketing automation, went viral, and reportedly made $15,000 in three days.
The pattern is consistent across multiple sources: the leverage available to a single person with access to agent infrastructure has crossed a threshold. You don't need a team to ship. You need the right agent stack and the judgment to know when it's working.
Perplexity's AI Computer Ships Inside Microsoft Teams
While the agent infrastructure race plays out at the infrastructure layer, the product layer continues to compress. Perplexity's general purpose AI computer — a digital worker that can conduct research, build dashboards, and draft documents — is now available directly in Microsoft Teams.
Here's everything you need to know:
- Available through Microsoft Marketplace and requires Pro, Max, or Enterprise account
- Embeds directly into Teams workspaces — no context switching required
- The dynamic: AI assistants are no longer competing on capability alone; distribution inside existing workflows is becoming the differentiator
- OpenClaw 5.3 released new features including file transfer for paired nodes and improved live agent control
- /steer command allows nudging active agent runs without derailing sessions — a small UX improvement that matters for agent reliability in production
⚡ Quick Hits
- Meta: ~$13B El Paso, TX AI data center (1 gigawatt, operational by 2028) — biggest single-site AI infrastructure bet in the company's history
- Apple: Lost App Store contempt appeal, now seeking Supreme Court review — App Store economics still in play
- China: New law prohibits companies from firing employees solely to replace them with AI — first major economy to codify this
- Amazon: Iranian drone strikes damaged UAE/Bahrain AWS data centers; months-long repairs underway
- Tesla: FSD crossed 10B miles engaged — but still requires human supervision, regulators haven't approved full driverless, Waymo already at 1M rides/week in 10 cities
- Fervo Energy: Geothermal startup planning Nasdaq IPO at ~$6.5B valuation, raising up to $1.3B
- SpaceX: Starship water-deluge test explosion puts May 12 launch in doubt
Techlook — AI & tech signal for founders and builders.