Microsoft Cuts Copilot as Bessemer Pegs AI Margins at 50-60%

Category Formation in One Week

Four independent actors converged on agent-payments infrastructure in the same week, which is usually the moment a category stops being a thesis and starts being something you can actually buy. Google Cloud and Solana shipped pay.sh, metering Gemini, BigQuery, and Vertex AI in stablecoins at $0.001 to $20 per call, with 75 providers live at launch and MCP-server compatibility. Anchorage Digital launched Agentic Banking with regulated settlement and Know-Your-Agent identity standards. Stripe shipped 280+ features including agent wallets and stablecoin rails. a16z crypto closed Fund 5 at $2.2B and named autonomous agent transactions as a deployment target, in writing.

When four independent actors converge on one category in one week, the integration problems that killed prior narratives just got retired simultaneously. That does not happen often enough to ignore.

The Comp Set Just Got Printed

AI-finance software valuations now have anchors, or rather, the anchors have decided to publish themselves: Rogo at $2B on a $160M Series D for IB modeling and research, 9fin at $1.3B on a $170M Series C for debt markets intelligence, and Mercury at $650M ARR, profitable for 3 years. These are almost certainly sub-$30M ARR businesses priced at 50-80x forward. Separately, Anthropic co-founded an AI-native services firm with Blackstone, Goldman Sachs, and Hellman & Friedman whose explicit job is deploying Claude across hundreds of PE portfolio companies.

Where the Alpha Actually Sits

The headlines favor the mega-names. The alpha sits one floor down:

1. Agent wallets, KYA identity, and agent orchestration at Seed/A. Underfunded relative to the demand curve pay.sh and Anchorage just created. Expect 3-5x multiple expansion once the first marquee Series A prints.
2. MCP-native data and guardrail plays in finance verticals. NatureAlpha's pattern — proprietary dataset, MCP server, monthly Claude Skills updates — travels into credit, private markets, alt-data, and regulatory reporting.
3. Agent-fraud detection and spend controls. Stripe just commoditized the payment rail. Value migrates one layer up, to which agent is transacting, what it is allowed to do, and whether the transaction is fraudulent.

The Timing Constraint

a16z's $2.2B with an explicit agent-payments thesis will reset Series A multiples in this wedge inside two to three quarters. Anything not written in the next two quarters gets written at 3-5x the price. The counter-thesis is worth stating in the same paragraph: fund size is not deployment velocity, a16z crypto has historically been patient, and if the marquee agent-payments Series A does not print by Q3 the window widens rather than closes. This is probably wrong on timing and approximately right on direction.

Finally, the 76% concentration on OpenAI across finance firms, paired with 43% of regulators not yet tracking AI adoption, is a latent regulatory catalyst sitting on the table. A single FSOC or OCC guidance note mandating model diversification triggers a category-wide repricing. The audit layer is cheap today.

AI Gross Margins Are Running 50-60%

Bessemer finally put the numbers in public, and they are worse than what most board decks will admit: AI companies are landing at fifty to sixty percent gross margins against the eighty to ninety percent that SaaS investors spent a decade treating as natural law. The fast-growth cohort runs closer to twenty-five percent gross margin and funds the gap with burn. OpenAI disclosed a thousand-fold response cost reduction over fourteen months, which sounds like the end of the problem until you notice reasoning models push per-query compute up by ten times and eat most of it back. The margin trajectory of any AI company now depends on whether its optimization stack runs faster than its reasoning workload mix grows.

A company doing $100M ARR at 25% gross margin is not the same asset as one doing $100M at 60%, and the gap between those two companies inside the same portfolio is about to widen considerably.

Microsoft Just Proved the Thesis

The best-resourced AI distributor on earth, with the cheapest compute and a captive OpenAI relationship, still cannot make horizontal AI-in-every-app pencil out. The Xbox CEO killed Gaming Copilot. The Windows chief pulled Copilot from Photos, Widgets, and Notepad. An EVP conceded on the record that inference costs are pressing on margins. What Microsoft is defending instead is Office 365 Copilot at $30/seat with 33% QoQ paid user growth. The read is simple: workflow-embedded AI with measurable ROI survives, feature-spray AI does not.

The sleeper detail is that Anthropic is now inside Microsoft's flagship AI revenue product, which quietly ends the OpenAI exclusivity narrative and raises Anthropic's strategic optionality by an amount the last round did not price.

Anthropic Moves to Per-Result Pricing

Anthropic is shifting from per-token to per-result billing, which is probably the most important unit-economics signal of the quarter. It structurally rewards vertical agent companies with measurable outcomes, and it structurally destroys wrapper businesses whose margins evaporate the moment the lab commoditizes the workflow underneath them. The counter-thesis, which I half-believe: token-based pricing is currently a structural moat for Anthropic and OpenAI, because customer gaming (Meta's engineers burning millions of tokens on internal leaderboards being the canonical example) actually increases lab revenue. Both things are true at once, which is why the transition matters.

The Repricing Framework

The practical number is this: a thirty percent inference cost reduction on a $70M run-rate drops $20M+ straight to EBITDA, which is a higher-ROI use of operating capital than most GTM line items currently on the same board's roadmap. The portcos that have already elevated inference ownership to a VP line with quarterly cost-per-query targets will separate from the ones still treating it as a platform problem.