AnthropicRents220KGPUsFromxAIasItTopsOpenAI

◆ DEEP DIVES

1. 01

   The xAI Lease Changes Everything: Enterprise AI Is Now a Duopoly With One Side Retreating

   What Actually Happened

   Anthropic leased the entire Colossus 1 cluster — 220,000+ NVIDIA GPUs including GB200s — from xAI, whose CEO has on the record called the counterparty 'misanthropic and evil.' Treat this as the puzzle it is. A frontier lab is renting roughly 45% of a competitor's current compute because the alternative was turning enterprise customers away while demand grew 80x against a 10x plan. That is not procurement. That is triage with a press release.

   Now set it against Ramp's April data: Anthropic at 34.4% of paid business share vs OpenAI at 32.3%. A 2.1-point lead after Anthropic quadrupled year-over-year and OpenAI grew 0.3%. The enterprise crown moved on spending data rather than survey vibes. That is the more interesting version.

   ---

   The Three-Body Problem

   Three signals landed inside the same week and they interact:

   • xAI is retreating from the frontier. Leasing 45% of your compute to a declared enemy is a concession dressed as a partnership. Grok has no visible path to B2B or B2C traction while DeepSeek and Qwen take developer mindshare. Reprice xAI exposure as infrastructure plus X-distribution, not a frontier lab.
   • Anthropic is capacity-constrained and winning. The silent Claude Code nerfs, the mid-trial Pro revocations, the corporate bans — none of it was strategy. It was supply rationing. Demand is the easy part.
   • OpenAI is defending, not expanding. A two-month free Codex enterprise promo within days of the Ramp flip is the move of a vendor that just lost a procurement review. Stack it with Altman's $2B in cross-holdings surfacing in court and the governance noise, and procurement committees now have filing-grade cover to multi-vendor.

   When rivals rent compute to declared enemies, you are not in a glut. You are in a shortage that bends strategy.

   ---

   Anthropic's Pre-IPO Margin Move

   On June 15, Anthropic converts every Claude subscription into a dollar-matched API credit pool, ending the 70-90% arbitrage that Cline, OpenCode and the rest of the third-party harness crowd were quietly running. This is margin recovery timed to IPO diligence, with a new CFO hire and a likely October listing as the timeline. Any portfolio company whose COGS assumed subsidized subscription tokens just lost 20-40% of effective runway, which is the sort of detail one prefers to learn before the term sheet, not after.

   What Sources Disagree On

   The honest question is whether Anthropic is taking share from OpenAI or simply expanding the pie. The Ramp data skews to SMB and mid-market via credit-card billing, and almost certainly understates OpenAI's 8-9 figure invoice-based enterprise contracts. The directional signal is real. The magnitude probably overstates the flip at the largest accounts. This is probably wrong, but: trim OpenAI secondary on the directional read, not the headline number.

   Action items

   • Audit every portfolio company's model-provider dependency within 48 hours — flag any with >50% token volume on a single provider and mandate multi-model routing roadmap
   • Re-underwrite any xAI/Grok exposure (direct secondary, SPV, or Grok-dependent apps) as infrastructure play, not frontier lab, using neocloud comps not Anthropic/OpenAI comps
   • Stress-test Claude-dependent portfolio companies' gross margins assuming June 15 credit unbundling eliminates the subscription arbitrage — request updated cohort data by month-end
   • Bid Anthropic secondary at sub-$700B before October IPO book-building firms up pricing — the enterprise share lead plus capacity constraints justify premium positioning

   Sources:The Pragmatic Engineer · TLDR AI · StrictlyVC · AINews · Morning Brew · The Hustle

2. 02

   Enterprise AI ARR Is Not SaaS ARR — The Observability Gap Creates a Category

   The ServiceNow Wake-Up Call

   ServiceNow, which by most measures is the most sophisticated enterprise software buyer alive, exhausted its full-year Anthropic budget by May 2026. Not because Claude underdelivered. Because Anthropic ships no granular per-user telemetry, no SLAs, and no enterprise dashboard that would have cleared procurement at a mid-tier SaaS vendor in 2014, and the National Life Group CIO said the quiet part on the record: Anthropic is 'great for consumer usage but not great for companies.'

   This is the company being valued at $900B+ on the premise that enterprise revenue carries the number. Consumer-grade plumbing does not usually carry enterprise-grade multiples.

   ---

   The FDE Consensus Validates the Problem

   Four firms independently arrived at the same conclusion, which is that the margin sits in deployment services rather than in the model:

   • Google Cloud is hiring hundreds of forward-deployed engineers
   • OpenAI stood up DeployCo with Bain Capital, acquiring a consulting firm for its 150-FDE starting roster
   • Salesforce and ServiceNow are staffing the same function in parallel

   When the industry concedes what Palantir figured out twenty years ago — that deployment is the bottleneck — the interesting question is who builds the software that eliminates 60% of FDE labor, because that company captures margin the labs structurally cannot defend.

   Enterprise AI spend reverses quickly once cost-efficient alternatives land. The CIO sentiment is the kind that historically precedes RFP cycles, not renewal cycles.

   ---

   The Investable Layer

   The AI observability and FinOps category is forming in real time, with three proof points worth naming:

   • ServiceNow's AI Control Tower, selling token-cost management to the same customers currently panicking about Anthropic bills
   • Modal at $4.5B, inference orchestration capturing the compute-management wedge
   • 'Tokenmaxxing' entering the enterprise vocabulary, with managers at Microsoft and Amazon coaching reports to inflate AI usage metrics, which means reported adoption is partially performative

   The pattern is recognizable, or rather the more honest version of it is: when CDIOs are simultaneously buying and building the same product, a Datadog-scale category is forming, and no independent winner exists yet, which is what keeps the 6-12 month sourcing window open.

   Revenue Quality Implications

   For anyone holding LLM-layer or AI-wrapper positions, the reversibility discount belongs in the marks. Enterprise AI spend with no SLAs, no telemetry, and no contractual lock-in carries a cliff-shaped risk profile that standard ARR multiples do not price, so apply a 20-40% haircut where those features are absent. This is probably wrong in detail, but the framework is honest: subscription revenue with the retention characteristics of consumption revenue and the pricing assumptions of enterprise SaaS.

   Action items

   • Build a sourcing sprint on AI observability/FinOps companies (token-cost attribution, per-user spend caps, SLA monitoring across model APIs) — target 5 meetings within 30 days
   • Demand SLA and usage-telemetry roadmap documentation from every model-layer company in portfolio claiming enterprise ARR — surface in next board cycle
   • Apply a 'reversibility discount' of 20-40% to any AI portfolio company's ARR multiple where SLAs, telemetry, and contractual lock-in are absent
   • Source Palantir-alumni founders for deployment-services software companies — they're the target talent pool for every FDE hiring wave

   Sources:Laura Bratton · Martin Peers · The Pragmatic Engineer · Bloomberg Technology · The Information AM · TLDR AI

3. 03

   Platform Owners Are Closing the Agent Window — Source Now or Lose the Layer

   The Convergence Signal

   One week, five moves, and the trick is that they line up: SAP committed €100M to an Autonomous Enterprise partner fund, Google embedded autonomous task execution into Android at the OS level (Gemini Intelligence, shipping summer 2026 on Galaxy S26 and Pixel 10), Apple began building agent governance into the App Store so agents cannot quietly spin up unapproved sub-apps, Notion launched a developer platform hosting Claude, Codex, Cursor, and Devin as teammates, and Airtable deployed $10M in credits through Hyperagent to five hundred founders.

   This is not a product week. It is category formation being drawn by incumbents before pure-plays can name it, which is the opposite shape from the cloud transition, where AWS defined the layer and the incumbents arrived late and embarrassed. This time the platforms are first.

   ---

   The Vercel Production Data

   Vercel's first AI Gateway index, across more than two hundred thousand teams, is the cleanest production read available:

   • 59% of token volume is now agentic — agents are the majority case, not the edge case
   • Anthropic takes 61% of spend, mostly expensive Opus reasoning calls
   • Google takes 38% of volume, mostly cheap Flash throughput

   Two different businesses are sitting inside what we keep calling foundation models. The spend-versus-volume split makes multi-model routing the enterprise default and quietly closes the argument that one lab runs the table.

   When four platform owners simultaneously write nine-figure checks to own the agent layer, the pure-play window closes faster than most pipeline decks assume.

   ---

   a16z's GTM Thesis: Timing Is the Risk

   a16z argued in public that the majority of next-decade GTM enterprise value migrates from the system of record (Salesforce at one hundred forty billion, HubSpot at nine billion) to a system of intelligence layer, and the Stitch check is the thesis with money behind it. Lemkin's proof point is cleaner than the slogan: ten-plus human seats collapsed to two plus one API seat, while spend rose 83% ($12K→$22K) with twenty-plus agents running. Seats down. Bill up.

   The counter-thesis deserves airtime. Salesforce absorbs the intelligence layer by acquisition within eighteen months, the wedge narrows, and the venture math gets uglier. This is probably wrong, but only because Lemkin says startups are moving faster than Salesforce and ServiceNow on agentic fabric, which is the entire reason any of this is investable.

   Apple's Walled Agent Garden

   Apple building agent governance into the App Store is the toll road that reprices every agent-layer bet that assumed iOS distribution was free. A WWDC reveal is plausible within weeks, which means any portfolio company whose GTM assumes friction-free agent install or non-App Store monetization on iOS needs a contingency before the keynote. After the announcement, the repricing is public and the conversation is worth less.

   Action items

   • Stress-test every mobile AI agent portfolio company with an 'Android OS parity' checklist — if Gemini Intelligence ships the feature in summer 2026, does the business still exist? Require written answer by June 1
   • Source 3-5 agent infrastructure deals in MCP governance, agent identity, and agent observability at Seed/Series A pricing this quarter
   • Map pipeline against 'agent-hosting platform' displacement risk — identify which deals get killed if Notion, Airtable, or Cursor absorbs the workflow
   • Rerank AI-GTM pipeline by orchestration moat depth — prioritize narrow high-frequency workflows with measurable outputs over horizontal 'AI copilot for sales' plays

   Sources:a16z · Simplifying AI · TLDR IT · ben's bites · Techpresso · TLDR

4. 04

   AI Security Left the Lab: LiteLLM on KEV, Daybreak's Partner Trap, and the 10x Cost War

   Three Production-Grade Signals in One Week

   The AI security category crossed from thesis to budget line in a single cycle:

   1. LiteLLM hit CISA's Known Exploited Vulnerabilities catalog — the first LLM-routing control plane federally flagged as actively exploited. This is the 'MongoDB ransomware moment' for AI infrastructure.
   2. Anthropic's Mythos became the first model to clear both UK AISI simulated attack ranges, with Congress routing access through NSA (not CISA) — signaling offensive/IC-led procurement, not civilian distribution.
   3. Microsoft's MDASH shipped 16 validated Windows CVEs in a single Patch Tuesday — first production evidence autonomous vulnerability discovery works at hyperscaler scale.

   ---

   The DepthFirst vs. Mythos Cost War

   DepthFirst's Open Defense Initiative landed FFmpeg, Envoy, and Kata as anchor tenants and published a pointed benchmark: 12 memory corruption bugs for ~$1,000 in compute vs. Anthropic's Mythos missing the same bugs across hundreds of scans at ~$10,000. A claimed 10x cost advantage using specialized harnesses rather than frontier-model rental.

   The bifurcation matters: a cost-led challenger and a capability-led incumbent produce a market where the middle gets squeezed within quarters. The interesting businesses to own are either the very cheap one or the very expensive one.

   Cyber-AI is pricing as an AI feature and transacting as a defense-tech category. The Series B window closes on whichever framing the market agrees to first.

   ---

   OpenAI Daybreak: The Partner Trap

   OpenAI launched Daybreak with Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Zscaler, Akamai, and Fortinet as 'partners.' This is the setup pattern — identical to prior platform cycles — of a platform preparing to disintermediate its partners. The partners know this. They signed anyway, because the alternative was being the one vendor not in the announcement. Cloudflare cutting 20% (1,100 people) citing 'the agentic AI era' in the same week suggests at least one name already reads the wind.

   Meanwhile, TrustedSec demonstrated that LLMs can reverse-engineer all five major commercial EDRs in days (vs. weeks previously), revealing they all share identical architectural furniture — YARA rules, Lua engines, local ML classifiers. The detection IP moat is eroding from both sides.

   The Portfolio Implication

   Any security portfolio company whose value prop assumes human-speed attacker cadence is a repricing candidate. Any whose deck assumes detection-rule IP is durable faces compression from AI-assisted reverse engineering above and platform absorption below. The survivors have distribution, data scale, or harness-level orchestration IP that's expensive to replicate per target.

   Action items

   • Run portfolio-wide exposure sweep for LiteLLM, Ollama, and PraisonAI usage within 72 hours — any portco running affected versions of these AI infrastructure components needs patched-or-mitigated confirmation
   • Request DepthFirst data room and validate the 10x cost claim against Mythos on 2-3 additional codebases before their round reprices
   • Build target list of 5-8 AI-native identity/deepfake defense companies at Series A/B — initiate conversations before the $40B 2027 TAM number gets absorbed into consensus
   • Add OpenAI Daybreak to competitive tracker for every security portfolio company — require each founder to articulate why their product isn't a Daybreak connector in 18 months

   Sources:CyberScoop · TLDR InfoSec · Clint Gibler · The Hacker News · SANS AtRisk · The Information AM