The Investor

The Two-Sided Squeeze

On May 12-13, Anthropic converted every Claude subscription into a dollar-matched API credit pool, which is a polite way of saying that a $200 plan now buys exactly $200 of programmatic tokens and not a token more. The seventy to ninety percent arbitrage that third-party harnesses (Cline, OpenCode, and a long list of portfolio-company dev tools) had been quietly running against subscription tiers is gone. OpenAI answered inside the same news cycle with two months of free Codex for enterprise switchers.

The timing is not an accident, or rather, the more interesting version is that it lines up with a new CFO and a likely October IPO. Anthropic is cleaning up margins for public-market diligence while defending the Ramp lead — 34.4% versus OpenAI's 32.3% in business spend, the first documented leadership flip.

Any Claude-dependent portfolio company whose COGS ran against subscription tokens has lost 20-40% of runway since Friday. The change is four days old — most founders haven't flagged it yet.

The Revenue Quality Problem

ServiceNow — one of the most sophisticated enterprise buyers on earth — exhausted its full-year Anthropic budget by May, which tells you something about either the appetite or the meter. Anthropic offers no granular per-user telemetry, no SLAs worth the name, and no enterprise dashboard that wouldn't embarrass a mid-tier SaaS vendor circa 2014. National Life Group's CIO put it without ornament: Anthropic is great for consumer usage but not great for companies.

This is the structural contradiction at the $900B valuation mark. The revenue justifying that number has consumer-grade stickiness: no contractual lock-in, no switching costs beyond re-prompting, and budget overruns that procurement will eventually catch. Secondaries are pricing in SaaS-grade retention. The evidence is not that.

Where the Alpha Sits

This is probably wrong, but multiple sources converge on the same read: the gap between Anthropic's model quality and its enterprise infrastructure creates a standalone category. ServiceNow is already selling AI Control Tower into the gap. Google, OpenAI (via Bain's DeployCo), and Salesforce are hiring hundreds of forward-deployed engineers. The Palantir playbook, where deployment services capture more margin than the product, is now consensus.

The Majority Case

Vercel's first production AI Gateway index, drawn from downstream data across more than 200,000 teams rather than lab benchmarks, puts agentic workloads at 59% of all token volume. So the agent era is not the upside case anymore. It is the base rate, in production, today. The interesting bit is the split underneath: Anthropic takes 61% of spend on the expensive Opus reasoning calls while Google takes 38% of volume on cheap Flash throughput. Two entirely different businesses are emerging inside the thing we keep calling 'foundation models.'

Multi-model routing is now the enterprise default, which means switching costs at the model layer approach zero. The moat has moved one layer up — to orchestration, workflow data, and institutional context.

Platform Owners Are Moving

Four platform moves landed inside seven days, which is either coincidence or, more likely, incumbents absorbing the agent layer before the pure-plays can define what it is.

• Apple — building agent governance into the App Store; gating agents through review and fee capture; possible WWDC reveal in weeks
• SAP — €100M Autonomous Enterprise fund, wiring NVIDIA and Microsoft into the platform layer
• Notion — developer platform with Claude and Codex as hosted teammates; External Agents API commoditizing the harness layer
• Google — Gemini Intelligence embeds autonomous task execution directly into Android at 97%+ share in emerging markets

a16z formalized the thesis in the same week: value migrates from system of record to system of intelligence, with the Stitch investment as the deployed check. Jason Lemkin's number is the cleanest unit-economics template anyone has put on the table — cutting from ten-plus human seats to two humans plus twenty agents while spend rose 83%, from twelve thousand to twenty-two thousand dollars. The customer paid more and was happier about it.

The Investable Window

This is probably wrong, but the window looks specific: 12-18 months before incumbent absorption compresses returns. The wedges that survive are narrow and vertical, not horizontal, and they are not the ones with the best demos:

The counter-thesis, which is not crazy: Salesforce, ServiceNow and SAP absorb the orchestration layer in one quarter rather than four, and the window closes before the checks clear. Enterprise software has done this before. The SAP fund announcement makes it more likely, not less.

Rivals Renting From Enemies

Anthropic leased the entire Colossus 1 cluster — 220,000+ NVIDIA GPUs including GB200s — from Elon Musk's xAI, which is interesting because the same Musk has publicly called Anthropic 'misanthropic and evil,' and people who think you are evil do not generally rent you their best asset unless the alternative is worse. This is not a partnership. It is what a shortage looks like when demand outruns plan by a factor most plans cannot survive: Anthropic modeled 10x growth and got 80x. The silent Claude Code nerfs, the Pro revocations, the enterprise bans — capacity triage dressed up as product roadmap.

When rivals rent compute from declared enemies, you are not in a glut. You are in a shortage that bends strategy. The 'AI compute glut' is a public-markets narrative, not a private reality.

The Neocloud Reality Check

Nebius gave us the cleanest public read on neocloud unit economics anyone has bothered to publish: $2.47B capex against $2.26B operating cash flow in Q1, revenue up 684 percent, four-plus customers bidding per GPU. The business is structurally cash-negative at scale. That matters for anyone marking private positions to comp:

The xAI Repricing

Several independent reads converge on the same conclusion, which is rare enough to take seriously: xAI is no longer operating as a frontier lab. Leasing roughly 45 percent of current compute to a direct competitor, trailing DeepSeek and Qwen in developer surveys, and showing no B2B or B2C product traction puts the correct comp set at neoclouds plus X-distribution, not OpenAI or Anthropic. Anyone offering xAI at frontier-lab multiples in a secondary is mis-marked.

The counter-thesis, and it deserves a hearing, is that xAI keeps the Grok roadmap and uses lease revenue to fund the next training run, at which point this was financing, not concession. That reading needs you to ignore the talent walking out the door (Cursor is already buying xAI engineers) and the benchmark lag, which is a lot to ignore.

Where Alpha Persists

The scarcity thesis holds. The margin structure underneath it is ugly. Alpha has migrated off GPU neoclouds, where the cash burn argues for multiple compression, into the adjacent layers: Nvidia ecosystem (networking, optics, liquid cooling), energy-for-AI with signed hyperscaler offtake, and inference silicon with model-lab design wins. The picks-and-shovels trade, or rather the more interesting version of it.

The Legitimation Event

LiteLLM, an AI gateway routing layer that quietly sits inside a lot of enterprise AI deployments, just landed on CISA's Known Exploited Vulnerabilities catalog. This is the first time AI-specific infrastructure has been federally flagged as actively exploited. The interesting part is not the bug. It is that 'AI security' stops being a pitch slide and becomes a procurement line with regulatory backing, which is a different sales motion entirely. Add Ollama's GGUF model-loader bug (CVSS 9.1, data exfiltration via malicious model files) and six critical CVEs in OpenClaw, and the AI-infra stack is roughly where enterprise SaaS sat in 2014, security-maturity-wise. That comparison flatters nobody.

In the same week, Microsoft's MDASH system shipped sixteen validated Windows vulnerabilities in a single Patch Tuesday, which is the first production evidence (rather than the usual demo) that multi-model AI vulnerability discovery actually works at scale. PraisonAI's auth-bypass CVE was weaponized within four hours of disclosure. AI-agent frameworks are now active attacker targets, and the discovery side is no longer theoretical either.

Platform Disintermediation Begins

OpenAI launched Daybreak with Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Zscaler, Akamai, and Fortinet listed as 'partners.' This is the classic platform-eats-partners setup, or rather the version of it where the partners know exactly what is happening and show up anyway because not showing up is worse. In the same week, TrustedSec showed that LLMs can reverse-engineer all five major commercial EDRs in days, pulling YARA rules, scoring thresholds, and exclusion lists out of products whose entire moat was the opacity of their detection logic. The moat was opacity. Opacity does not survive contact with a sufficiently patient model.

The Bifurcation Worth Funding

DepthFirst's Open Defense Initiative claims ten times the cost efficiency of Anthropic's Mythos on vulnerability discovery, finding twelve memory corruption bugs in FFmpeg for roughly a thousand dollars while Mythos missed them across hundreds of scans at around ten thousand. This is probably wrong in detail, but the shape is right: a specialized-harness-versus-frontier-model split that supports two viable businesses, very cheap or very expensive. The middle, as usual, gets squeezed.

The Mythos angle has a separate thesis. Clearing both UK AISI attack ranges and earning NSA access (not CISA) routes Anthropic toward IC-led procurement, which prices and transacts on a completely different curve. Cyber-AI is being priced as an AI feature and sold as a defense-tech category, and the Series A/B window closes on whichever framing the market agrees to first. The market has not decided yet.