The Investor

The Structural Shift

In five trading days, Wall Street built a distribution channel for AI that did not exist ninety days ago. A nineteen-firm consortium led by Blackstone committed ten billion dollars to push OpenAI through consortium-owned businesses. Five days later Anthropic closed a $1.5B JV with Blackstone, Goldman, Hellman & Friedman, and General Atlantic using the identical template. These are not funding rounds, or rather they are not only funding rounds. They are channel capital — a mandate to embed foundation-model AI into several thousand mid-market companies by operating-partner decree.

The competitive read is in the ratio. OpenAI took 6.7x the capital for roughly the same portfolio reach. Anthropic bought distribution parity at a discount, which looks rational once you read the Uber data point below on Claude Code's revenue quality problem.

A single PE sponsor decision now deploys an AI stack across 250+ portfolio businesses. The consortium footprint runs into the low thousands. This is the fastest new enterprise distribution channel built in a decade.

What This Means for Your Pipeline

Any AI startup selling into PE-owned mid-market now meets a gatekeeper at the door. If the sponsor is in the consortium, the default answer to "which AI stack?" is the lab partner, and independent vendors either win sponsor-level approval or get routed through the consortium's AI layer. Call it a 'Bessemer vs. Salesforce AppExchange' dynamic, except the marketplace is the entire PE mid-market and the platform owner is a frontier lab.

Who wins and who loses

The counter-thesis deserves airtime, and this is probably wrong, but: PE portfolio companies will revisit these AI choices in 18 months when contracts renew, and the JVs bought distribution rather than loyalty. Eighteen months of captive deployment still creates switching costs that independent vendors will find expensive to displace.

The Adjacent Cost Signal

Underneath the distribution story, Uber burned its full 2026 AI coding budget in four months on Claude Code at $500–$2,000 per engineer per month. The 'AI tools are cheaper than the engineers they augment' thesis, which underwrites most dev-tools pitches currently in pipeline, is empirically cracking. Into that cost crisis, IBM Granite 4.1 shipped at 30B parameters, 512K context, Apache 2.0. Zero licensing cost becomes a real substitution threshold for the first time.

Distribution gets harder at the top (the PE gatekeeper) and pricing gets floored at the bottom (open weights). API-wrapper business models face bidirectional pressure that most current marks do not reflect.

The Category Birth

Jack Clark, Anthropic co-founder and one of the more calibrated forecasters on this beat (the bar is low, he clears it anyway), is now putting 60%+ probability on fully automated AI R&D by end of 2028, with thirty percent by end of 2027. This is not a vibes number. He builds it from benchmark data showing every capability required for AI-building-AI sitting on a twelve to eighteen month saturation curve.

The numbers do the arguing, so let them:

• SWE-Bench: two percent on Claude 2 in late 2023, 93.9% on Claude Mythos Preview. Two years.
• METR task horizon: thirty seconds on GPT-3.5 in 2022, twelve hours on Opus 4.6 in 2026. Roughly ten times per year, which is the sort of slope one usually only sees on a pitch deck, not a benchmark.
• Training optimization: AI-driven speedups went from 2.9x in May 2025 to 52x in April 2026, against about 4x for human researchers given four to eight hours.

AI systems are now meaningfully superhuman at optimizing the code used to train their successors. That is no longer speculative. It is a measured capability with a published benchmark.

Capital Is Already Forming

Recursive Superintelligence raised $500M explicitly to automate AI research, the first pure-play at venture scale. Mirendil exists with the same charter, or rather the same pitch deck. OpenAI committed to shipping an 'automated AI research intern' by September 2026. The interesting part is not whether the intern ships on schedule (it will not) but that the spend is being underwritten as if it will.

The ICLR 2026 workshop on recursive self-learning, plus Karpathy's autoresearch work, give the category technical legitimacy. Call it the first post-scaling-laws AI thesis with real capital behind it.

What this means for the book

Every AI infra or tooling deal currently in pipeline needs to be re-underwritten against a world where the buyer itself may be partially automated within 24 months. That is a different deal than the one being pitched. Specifically:

The compounding error problem is the under-priced risk here: a 99.9% accurate alignment technique degrades to 60.5% after 500 recursive generations. That produces a regulated-market spend category (eval, alignment, interpretability) within eighteen to twenty-four months, assuming regulators move, which they usually do after the second incident rather than the first.

Where the Alpha Sits

Probably wrong, but: the alpha is not in picking which frontier lab wins the recursion race. That is priced. It sits in three adjacent layers, and the counter-thesis is named on each because I have been burned on exactly this shape of call before.

1. Kernel and training-stack optimization. The 52x speedup is the proof point. Cuda Agent, AscendCraft, DeepSeek's kernel work. Easily verifiable rewards, clear ROI, window closes in two to three quarters. Counter: the incumbents internalize this work instead of buying it.
2. Eval, interpretability, and alignment infrastructure. The compounding error problem builds a regulatory-driven spend category. Most peers still file it under 'research,' which is what they were doing with compliance software in 2013.
3. The mid-layer short. AI coding assistants without distribution, benchmark-only model startups, and the 'makes your senior engineer 2x productive' pitches are all about to see willingness-to-pay erode as the underlying talent gets automated. The market decides whether it cares.

The Scoreboard

Two numbers from the week point in opposite directions at the same category, which is the interesting part. Sierra raised roughly one billion dollars at a fifteen billion dollar valuation — seventy-five times trailing, about fifty times forward ARR — on outcome-based enterprise pricing, with ARR running from one hundred million to over one hundred and fifty million in three months. In the same week, a single GitHub Copilot session burned two hundred and twenty-one dollars of inference against a forty dollar monthly subscription. The top of applied AI is still bid aggressively. The business model underneath most of the category is quietly broken.

The seat-priced incumbents are funding the category's compute bill while the outcome-priced upstarts collect the category's margin. That is not a temporary arrangement. That is the arrangement.

The Cost Floor Is Collapsing From Below

Three data points converge. Deepclaude proxies Claude Code's UX onto DeepSeek V4 Pro at seventeen times lower cost, with an identical agent loop. Mistral Medium 3.5 shipped open-weights at 77.6% SWE-Bench, self-hostable on four GPUs. IBM Granite 4.1 at thirty billion parameters, five hundred and twelve thousand context, Apache 2.0, with zero licensing cost.

The pricing spread is now wide in a way that is not sustainable — Claude Pro at roughly ten times per token versus competing plans, Codex the most subsidized of the lot. When five vendors ship similar capability within a quarter of each other, the capability is not the moat. Distribution is, or workflow lock-in is, or nothing is, and the last option is on the table.

Where the value migrates

The harness and orchestration layer, rather than the model, is quietly becoming the place value is captured. The evidence:

• Prompt and middleware changes alone moved gpt-5.2-codex from 52.8% to 66.5% on Terminal-Bench, which is a model-sized improvement with no new model in it
• Open-model and good-harness combinations deliver more than twenty times cost reduction, per Vtrivedy's thesis
• Zyphra's folded TSP parallelism hit one hundred and seventy-three million tokens per second versus eighty-six million for standard approaches on 1024 MI300X GPUs, or a two times throughput gain on AMD

The defensible layer has moved to multi-model orchestration, evals, and governance. Each successive AI coding leader was built by a team ten to one hundred times smaller than the one it displaced, running Copilot to Cursor to Claude Code to OpenClaw, the last built solo. Four generations, no durable application-layer moat anywhere in the sequence.

The Sort

This is probably wrong in places, but the agent thesis does not need a rewrite so much as a sort:

1. Outcome-priced vertical agents — the Sierra and Harvey archetype, enterprise-grade, where fifty to seventy-five times ARR is justified by sequential growth at the hundred-million-plus mark. Long.
2. Harness and orchestration middleware — model-agnostic, context-pipeline moats, under-priced because the category has not been named yet. Source now.
3. AMD inference infrastructure — Zyphra's two times throughput gain is the first credible data point that AMD economics undercut NVIDIA for open-weight agent workloads at scale. Early, and real.
4. Flat-rate coding SaaS — every position here needs a gross-margin stress test against agentic load, and the counter-thesis is that Copilot renegotiates contracts before the arithmetic catches up. The two hundred and twenty-one dollar number is the mean once users unlock agentic workflows, not the outlier. Forced repricing within two quarters is the base case.