The Investor

The Strategic Concession

Anthropic has leased the entire Colossus 1 cluster, 220,000-plus NVIDIA GPUs including GB200s, from Elon Musk's xAI. That is roughly forty-five percent of xAI's compute, rented to the company Musk publicly called "misanthropic and evil." Renting compute to a declared enemy is what a shortage bending strategy looks like. It is not what equilibrium looks like.

Dario Amodei has now conceded that Anthropic planned for 10x growth and got 80x. The quiet Claude Code nerfs and mid-trial Pro revocations were not strategy. They were capacity triage. This is the demand signal every neocloud thesis has been pricing in, arriving with named counterparties attached.

xAI's Retreat Is the Under-Discussed Signal

Renting forty-five percent of your compute to a competitor is, whatever else one calls it, a concession in the frontier race. xAI has no meaningful B2B or B2C traction, trails DeepSeek and Qwen in developer surveys, and its parent just became a landlord. The honest comp set is now neoclouds plus X-distribution, not Anthropic or OpenAI.

Any fund or secondary offering xAI at frontier-lab comps is mis-marked. Grok has no revenue, lost developer mindshare, and its compute is generating more value as rental income than as training runs.

Neocloud Economics: Real Demand, Ugly Unit Economics

Nebius is the public benchmark, or rather the only public benchmark we have. 684% Q1 revenue growth on $399M, guiding to three to three-point-four billion for 2026, with four-plus customers bidding per GPU. The catch sits on the cash statement: $2.47B of capex against $2.26B in operating cash flow. Structurally cash-negative, growing extraordinarily.

This is probably wrong, but: private neocloud rounds at eight to ten times forward revenue should compress to two to three times once the market accepts that capex exceeds operating cash on a permanent basis. The demand is real. The unit economics are not venture-returnable at the current entry multiple unless the exit is a strategic acquisition.

The 'Compute Glut' Narrative Is Dead

Public-market bears citing overcapacity are trading against a private market in which frontier labs rent from sworn enemies. The gap between the two is the arbitrage. The trade is picking the right layer of the infra stack, not blanket exposure.

The ServiceNow Wake-Up Call

ServiceNow, which is as sophisticated an enterprise software buyer as exists on this planet, blew through its full-year Anthropic budget by May 2026. Not because Claude underdelivered. Because Anthropic provides no granular per-user or per-tool usage telemetry, no enterprise dashboards, and no SLAs worth the name. National Life Group's CIO put it without ornament: Anthropic is 'great for consumer usage but not great for companies.'

This is probably wrong, but: the market is being asked to value Anthropic at $900B+ on the premise that enterprise revenue justifies the number, and consumer-grade plumbing does not usually justify enterprise-grade multiples. The revenue is real. The quality of that revenue — or rather, its stickiness, its predictability, and whatever contractual protection sits underneath it — is a different question, and the answer is the one that matters.

The FDE Consensus: Deployment Is the Bottleneck

Four organizations independently concluded the margin lives in deployment services this quarter, which is the kind of coincidence that usually isn't:

• Google Cloud is hiring hundreds of forward-deployed engineers
• OpenAI/Bain stood up DeployCo and acquired a consulting firm for its 150-FDE roster
• Salesforce is staffing the same function
• ServiceNow built AI Control Tower to sell into the problem it discovered internally

When four firms independently conclude the margin is in deployment services rather than the model, the margin is probably in deployment services rather than the model.

The AI Observability Category Is Forming Now

ServiceNow's response is the instructive part: it built AI Control Tower and is now selling it to other customers panicking about their Anthropic bills. A CDIO watched the problem form inside her own P&L and decided to be the vendor rather than the victim, which is the Datadog playbook with the serial numbers barely filed off.

The category ingredients are all present: token-level cost attribution, per-user spend caps, SLA monitoring across model APIs, and budget governance. No independent winner exists yet. The 6-12 month sourcing window is open, and what the incumbents are not doing during that window is shipping the missing telemetry themselves.

What This Means for LLM-Layer Marks

Enterprise AI spend is reversible in ways traditional SaaS is not. Low switching costs, no contractual lock-in (because no SLAs), FOMO-driven procurement, and budget overruns discovered only at billing time produce a cliff-shaped risk profile rather than a curve. Apply a 20-40% 'reversibility discount' to any model-layer ARR multiple where SLAs and telemetry are absent. The thesis could be wrong if Anthropic ships enterprise controls in the next two quarters, or if procurement teams decide they don't care. Neither has happened yet.

59% Is the Base Rate, Not the Upside Case

Vercel published what it calls the first production-grade index of real AI usage across more than 200,000 teams, and the headline number is that agentic workloads now carry 59% of all token volume. That is not a leading indicator. It is the majority case in production, which makes the chat-completion era a minority revenue line and most of the slideware from eighteen months ago retrospectively wrong.

The more interesting cut is the spend/volume bifurcation. Anthropic captures sixty-one percent of dollars on premium reasoning via Opus. Google moves thirty-eight percent of volume on commodity throughput via Flash. Two different businesses are now legible inside what we have been calling 'foundation models,' which validates multi-model routing as the enterprise default and retires the single-model moat thesis.

The June 15 Margin Event

Anthropic converted every Claude subscription into a dollar-matched API credit pool. A two-hundred-dollar plan now buys exactly two hundred dollars of programmatic tokens, and the seventy-to-ninety percent arbitrage that Cline, OpenCode and dozens of similar harnesses were running against subscription pricing is gone. OpenAI answered in hours with two months of free Codex for enterprise switchers.

Combine that with Claude Code's /goal mode absorbing the autonomous coding category and Notion's External Agents API commoditizing the interface, and the result is a two-sided squeeze on anyone whose business model was thin-wrapping a frontier model.

Every Claude-dependent portco that priced COGS off subsidized subscription tokens lost twenty to forty percent of runway last Friday. The change is four days old. Most founders have not flagged it to their boards yet.

Consolidation Accelerating from Above

The platform layer is moving to own the agent surface, and the inventory looks like this:

The pattern is consistent. Incumbents are consolidating the autonomous-enterprise category before pure-play agent startups can define it. The cloud transition ran the other direction. This time ERPs and workflow platforms moved first, which means the obvious horizontal land grab is mostly closed.

Where the Alpha Still Lives

This is probably wrong in one or two of these wedges, but vertical beats horizontal in agent ops from here. The investable surface is agent governance (the gap SAP and Levie keep flagging), agent identity/auth (eighty-one percent bot-detection bypass is a market by itself), evaluation and observability (the Haiku-as-judge pattern is now canonical), and vertical-specific orchestration in regulated industries. The horizontal seat at this table is taken.

The Category Winner Proof Point

Abridge raised $550M in 2025 alone (two hundred fifty million early in the year, three hundred million in June at a $5.3B mark), and now services 250 large US health systems with 80M+ projected patient-clinician conversations this year across 28 languages and 50 specialties. The interesting part, or rather the part worth underwriting against, is the data moat that compounds with every patient visit at a rate no foundation model or new entrant can backfill after the fact.

Health systems, which are historically the slowest enterprise buyers on earth, compressed Abridge's release cycles from quarterly or biannual to monthly. That single observation rewrites the velocity assumption in every healthcare SaaS DCF built in the last five years.

The Three-Act Underwriting Framework

Abridge's structure validates a template that probably extends to other regulated verticals, with the usual caveat that templates tend to work until they don't:

1. Save Time (documentation) is the wedge that gets the initial sale to the CMIO.
2. Save Money (prior auth, billing) is the expansion that captures the CFO budget, which is where the actual dollars live.
3. Save Lives (clinical decision support) is long-dated optionality that the FDA's January 2026 CDS guidance may or may not unlock; treat it as a maybe.

Each act monetizes the same proprietary conversation asset against a different buyer persona in the same account. You sell once and expand against data you accumulated for free as a byproduct of the first product, which is the actual structural advantage.

The Non-Compete Posture Is the Key Insight

Abridge positioned itself as a 'clinical intelligence layer' rather than an EHR replacement, explicitly non-competing with Epic and Cerner, which is the single biggest existential risk for any healthcare AI company and the one Abridge declined to take. The EHR remains the system of record and Abridge sits on top of it, which is now table stakes for any vertical AI deal in regulated industries.

What I'd underwrite for in vertical AI deals: a per-use data flywheel competitors cannot access, at least two monetization vectors beyond the wedge, and an explicit non-compete posture with the dominant system-of-record.

Where the Adjacent Alpha Sits

The ambient documentation category is closed. Three adjacent chapters remain open, in roughly descending order of how much I'd pay to be early:

• Payer-side prior authorization, where Abridge attacks from the provider side and the payer counterparty is greenfield.
• Nursing and short-form clinical workflows, where a thirty-second visit is a different product entirely from a fifteen-minute physician encounter.
• Specialty verticals outside US large systems (veterinary, dental, behavioral health, international markets) that Abridge won't prioritize for 24+ months.