AnthropicEnds70%SubArbitrage;TokenTelemetryGapBites

◆ DEEP DIVES

1. 01

   The AI Observability Gap: Your Portfolio Is Spending Blind — Fund the Fix

   ServiceNow Proved the Category Exists

   ServiceNow employs thousands of engineers and runs enterprise procurement at a level few buyers can match, and it still exhausted its full-year Anthropic budget by May because Anthropic provides no per-user, per-tool telemetry 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 the lab the market wants to value at $900B+ on enterprise revenue quality.

   The gap is not Anthropic's alone. It is structural at the model layer. No frontier lab offers granular usage attribution or cost-center allocation, and the enterprise-grade dashboards on offer would embarrass a mid-tier SaaS vendor from 2014. Revenue growing at 120x is revenue growing without basic financial controls on the buyer side.

   ---

   The FDE Model Is Now Consensus, Which Is the Tell

   Four organizations independently concluded the margin lives in deployment, not model access:

   • Google Cloud: hiring hundreds of forward-deployed engineers
   • OpenAI/Bain Capital: stood up DeployCo, acquired a consulting firm for its 150-FDE starting roster
   • Salesforce: staffing the same function internally
   • ServiceNow: shipping AI Control Tower to the same customers panicking about their Anthropic bills

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

   The Palantir playbook of forward-deployed engineers embedding in customer workflows is now industry consensus. But the software that replaces 60% of FDE work (context ingestion, custom eval harnesses, workflow templates, token cost attribution) does not exist yet. That is a category waiting to be funded.

   ---

   Where the Alpha Sits

   Two distinct wedges are investable at different stages:

   Category	Stage	Comp	Window
   AI FinOps / Token Cost Attribution	Seed–Series A	Datadog, Cloudability	6–12 months
   Deployment Services Tooling	Series A–B	Palantir Foundry (productized)	12–18 months
   Enterprise AI Control Tower	ServiceNow building	ServiceNow ITSM	Incumbent advantage

   The thing worth noticing: ServiceNow is both the customer with the problem and the vendor building the solution. That is the setup ServiceNow itself rode in ITSM a decade ago, where a CIO who lived the pain decided to be the vendor rather than the line item. Modal at $4.5B is the infrastructure comp. The observability layer sitting above it has no comp yet because no winner has emerged.

   The counter-thesis: Anthropic or OpenAI build enterprise telemetry themselves within 6 months, collapsing the wedge. Possible, but unlikely ahead of an October IPO. Anthropic's priority right now is margin recovery, not platform extension.

   Action items

   • Launch a sourcing sprint on AI observability / FinOps-for-AI startups at Seed-to-Series A within 30 days
   • Demand SLA and usage-telemetry roadmap from every model-layer portco pitching enterprise ARR before next board cycle
   • Map Palantir alumni network as both founder candidates and advisory hires for deployment-services diligence
   • Attend or debrief The Information's May 20 AI Infrastructure event (OpenAI's Sachin Katti, Amp's Anj Midha)

   Sources:Laura Bratton newsletter · Martin Peers analysis · The Pragmatic Engineer deep-dive · The Information AM briefing · TLDR AI report

2. 02

   Anthropic's June 15 Credit Change: 30 Days to Audit Your Wrapper Book

   The Pricing Move That Kills a Business Model

   On May 12, Anthropic converted every Claude subscription into a dollar-matched API credit pool, which is to say a $200 plan now buys exactly $200 of programmatic tokens and not a token more. That kills the 70-90% arbitrage the third-party harnesses (Cline, OpenCode, OpenClaw) had been quietly running by routing heavy agentic workloads through consumer subscriptions, paying twenty dollars for what would have cost $150-200 at API rates. The arbitrage was the business model.

   Full effect lands June 15. OpenAI answered inside the hour with two months of free Codex for enterprise switchers, which tells you what they think the prize is. Hold this next to Ramp's April print, where Anthropic at 34.4% of business spend vs. OpenAI at 32.3% marked the first documented lead change, and the read is straightforward. This is margin recovery in a developer-benefit costume, timed to a likely October IPO.

   ---

   Who Gets Hurt, and How Fast

   Any portfolio company whose gross margins quietly assumed subsidized Claude tokens is repricing this month:

   Company Type	Exposure	Margin Impact	Action
   Claude wrapper / harness	Direct — COGS jumps 70-90%	-20 to -40 pts gross margin	Urgent board conversation
   Multi-model coding agent	Moderate — can route to alternatives	-5 to -15 pts if Claude-heavy	Pull token-mix data this week
   Agent-native startups on Hyperagent credits	Clock-based — credits expire	Deferred pain	Model post-credit economics

   The two-sided squeeze is the harder problem, or rather the more interesting version of it. Anthropic meters tokens from below; Notion's External Agents API, hosting Claude, Codex, Cursor, Decagon, Warp and Devin inside the same workspace, commoditizes the interface from above. Thin wrappers get compressed from both ends at once.

   ---

   What Survives

   The moat has migrated off model access. Three positions still defend:

   1. Proprietary workflow data — accumulated context the model doesn't have
   2. Open-source distribution — Cline's rebuilt SDK with agent teams and scheduled jobs
   3. Enterprise-grade bounded execution — Cursor's cloud agents with rollback, scoped egress, and isolated secrets

   Anything that doesn't clear one of those three bars is a pass. The subscription arbitrage was a subsidy, not a moat, and it just expired.

   The October framing is the part allocators should sit with. A new CFO, margin-recovery pricing and a credit conversion in the same quarter is what cleaning up economics for public-market scrutiny actually looks like. This is probably wrong, but secondary marks should firm before September, which means the window to position pre-book-building is closing rather than open. What Anthropic is not doing is subsidizing anyone else's gross margin.

   Action items

   • Request updated gross-margin cohort data from every Claude-dependent portco by end of month — specifically model COGS at API rates vs. prior subscription rates
   • Stress-test autonomous coding agent deals in pipeline against Claude Code's /goal + Auto Mode as free baseline — require a defensibility answer that isn't 'better prompts'
   • Size Anthropic secondary position before September book-building; bid aggressively at sub-$700B given $900B+ primary signals
   • Update AI app-layer thesis memo: downgrade thin-wrapper scorecards, upweight proprietary workflow data and OSS distribution moats

   Sources:AINews analysis · ben's bites gateway data · Daily Dose of DS technical report · TLDR AI report · Techpresso briefing

3. 03

   Cyber-AI Just Became a Federal Budget Line — The 18-Month Funding Window

   Mythos Cleared Both Ranges — And Went to NSA, Not CISA

   Anthropic's Mythos became the first model to clear both UK AISI simulated attack ranges, which pushes the 'AI cyber tasks doubling every few months' line further up and to the right than even the optimists were modeling six months ago. The detail worth lingering on is not the capability. It is that the Congressional briefings that followed routed access to the NSA rather than CISA. CISA would have meant defensive civilian procurement, slow money, GSA schedules, the usual. NSA means offensive and IC-led spending. Different buyer, different budget, different scale.

   Meanwhile OpenAI rolled out Daybreak with Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Zscaler, Akamai, and Fortinet listed as 'partners,' which is the word platforms always reach for in the quarters before they disintermediate the people they are partnering with.

   ---

   The 10x Economics Bifurcation

   Anthropic is pricing cyber-AI from the top. DepthFirst's Open Defense Initiative came at it from below and found twelve memory corruption bugs in FFmpeg for roughly a thousand dollars in compute, against Mythos missing the same bugs across several hundred scans at around ten thousand. A tenfold cost advantage, specialized harnesses over frontier-model rental.

   Capability-led from above, cost-led from below. The middle of this market gets squeezed in quarters, not years, and the squeeze looks like this:

   Position	Player	Moat	Buyer
   Capability-led (top)	Anthropic Mythos, OpenAI Daybreak	Benchmark lead + gov access	IC/DoD, Fortune 100 CISOs
   Cost-led (bottom)	DepthFirst, specialized harnesses	10x economics on narrow surfaces	OSS maintainers, mid-market
   Squeezed middle	Generic 'AI finds bugs' wrappers	None surviving	Repricing targets

   ---

   The EDR Moat Is Cracking in Parallel

   TrustedSec showed that LLMs can reverse-engineer all five major commercial EDRs in days rather than weeks. The detection-IP moats at CrowdStrike, Palo Alto, and SentinelOne, all of which still trade at premium multiples, rest on YARA rules, Lua engines, and local ML classifiers that are now programmatically extractable. Daybreak parking itself one layer above them as a 'partner' platform is the second compression force operating on the same names at the same time.

   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.

   The Investable Wedges

   1. AI-native identity defense — forty billion dollar projected 2027 loss TAM, liveness plus behavioral plus device graph
   2. AI-speed exposure management — continuous red-team, autonomous patching at machine cadence
   3. LLMjacking defense — shared attacker tooling and 113K monthly requests on a single honeypot, which is what a pre-category looks like before anyone calls it one
   4. IC-aligned dual-use teams — TS/SCI talent, Five Eyes partnerships, offensive-AI primitives wrapped in compliance

   Action items

   • Build a target list of 5-8 AI-native identity/deepfake defense companies at Series A/B and initiate conversations within 30 days
   • Commission a 2-week portfolio-wide review of any security portco whose value prop assumes human-speed attacker cadence
   • Request DepthFirst data room and validate 10x cost claim against Mythos with independent benchmark on 2-3 additional codebases
   • Add 'Daybreak displacement risk' question to every security portco board agenda — require a written answer on what product isn't a Daybreak connector in 18 months

   Sources:CyberScoop intelligence · TLDR InfoSec analysis · Clint Gibler security brief · SANS AtRisk report · The Hacker News analysis