Welcome to Issue 171 of The CTO Show Brief.
The Capital Stack Is Re-Sorting Itself
The week's deals describe a single motion. Capital is consolidating around frontier compute while the operating layer is being re-licensed, the labor base re-priced, and the distribution layer pried open. The signals below are not separate stories. They are the same restructuring viewed from different desks.
Eighty percent is now the number that matters.
Q1 2026 global venture cleared roughly $300 billion, and AI and its physical compute stack absorbed $242 billion of it. Four rounds, OpenAI at $852 billion post, Anthropic in negotiation near $900 billion pre, xAI's Series E, and Waymo's growth round, accounted for around 65% of global deployment. The remaining ecosystem is functionally a different asset class, with early-stage still growing 41% year-over-year on $41 billion deployed but inside a market where pricing power, talent, and compute access tilt sharply toward the four. Allocators reading this stack are no longer choosing between AI and not-AI. They are choosing between proximity to the concentration and exposure to its second-order effects.
The 18-month thesis any LP should be testing now sits one degree off the obvious.
The labor line and the capex line have crossed.
Tech layoffs passed 128,000 across 286 events year-to-date, with March alone removing nearly 50,000 roles, while the same companies are projected to spend $700 billion on AI infrastructure in 2026. Block cut 40% of headcount as a declared AI remake, Oracle removed 30,000 roles to free $8 to $10 billion in annual cash flow, and ZipRecruiter shows senior postings at 43.1% of the market against 7.4% entry-level. This is not a hiring cycle. It is a substitution of payroll into silicon, with the surviving engineering profile reshaping toward integration, orchestration, and AI-fluent senior ICs. Mustafa Suleyman's 12 to 18 month claim on white-collar automation is the public version of what the layoff ledger already shows in private.
The operator question is no longer when this lands, but which functions absorb the displaced budget first.
Microsoft just rewrote the tax on autonomy.
The May 1 launch of Microsoft 365 E7 at $99 per user, the first new enterprise tier since 2015, bundles Copilot, Agent 365, and Entra into a single line item priced to pull every enterprise into the premium AI stack. Agent 365 is the meaningful object, a control plane that treats AI agents as identity-bearing users inside the Entra graph at $15 per seat, while the rewritten multiplexing clause replaces "hardware or software" with "any method, including automation." Every human benefiting from an agent's connection to M365 now requires their own license, which closes the obvious arbitrage that agent-native startups were quietly counting on. The seat-based model is not dying. It is being extended to non-human seats.
The defensibility map for any agent-layer startup looks materially different after this clause than before it.
Apple's distribution layer is fragmenting in OpenAI's hands.
OpenAI has retained external counsel to weigh breach-of-contract action against Apple, while Apple has structured a multiyear Gemini deal at roughly $1 billion annually and is preparing iOS 27 at WWDC to support a multi-model marketplace including Claude and Gemini at the Siri permission layer. The original 2024 integration that was supposed to drive billions in ChatGPT Plus conversion never landed, and OpenAI is now routing around Apple through sovereign deals like the Malta national rollout and the Jony Ive hardware program. The platform layer that looked locked in 2024 is now a contested marketplace where model providers compete for placement rather than receive it. Distribution is becoming a procurement decision again, not a partnership.
The implication for any company whose distribution thesis assumed a single platform gatekeeper deserves its own re-underwriting.
Regulation is moving from policy to live stress test.
The federal executive branch has directed Justice to challenge state-level AI laws while Colorado has substantially rolled back its AI Act, producing a fractured domestic regime where sector-specific rules survive and comprehensive ones do not. Connecticut moved on employment AI, the administration is signaling alignment with Beijing's pre-release model review posture, and Brazil's 2026 presidential cycle has become the first real-world test of synthetic personas generated on Gemini and Flow operating at scale on hyper-local political content. The compliance perimeter is no longer a future engineering cost. It is a present procurement filter that determines which agent-layer products can be sold into regulated buyers at all.
The startups that priced governance as a feature are about to find out whether they priced it correctly.
Closing line: the next six months will resolve whether the agent economy gets built inside Microsoft's licensing graph, around it, or against it, and that single resolution will reprice most of the cap tables written in the last twelve months.
🎙️Episodes Recap:In this episode of The CTO Show with Mehmet, Mehmet sits down with Laura Fu , GTM Architect at DevRev. Laura brings a RevOps and sales enablement lens to a question many GTM leaders are now facing: AI does not fix sales by sitting on top of old workflows. The conversation reframes AI in go-to-market as an operating model problem, not a tooling problem. Laura argues that AI-native execution requires new feedback loops, better data capture, agent-readable systems, and a different view of enablement. The strongest claim is that dashboards and forecast calls become less central when agents can surface the signal directly.
In this episode of The CTO Show with Mehmet, Mehmet sits down with Omid Pakseresht , CEO of Goodfolio. Omid works on enterprise AI systems that move beyond pilots and into real business workflows. The conversation reframes enterprise AI failure as a systems problem, not a model problem. Omid argues that most AI initiatives break because the workflow, ownership model, governance layer, audit trail, and adoption path were never designed properly. The model may work, but the enterprise system around it often does not.
📖 From Nowhere to NextThanks for reading — and for being part of this growing, global-minded network.
— Mehmet
