Are we in the Agentic Inflection point or not?

Yes, this is a private markets newsletter, but today's news is mostly about public companies. We'll get back to private companies — promise.

If you read the analyst notes on agentic AI, you'll notice: the headline skeptical numbers haven't been refreshed.

Gartner, June 2025: over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Still the most-cited stat in every analyst briefing eleven months later.

MIT, July 2025 — "The GenAI Divide": 95% of enterprise AI pilots deliver no measurable P&L impact. Still anchors the bear case.

McKinsey, 2025: 62% of enterprises are experimenting with agents, but fewer than 25% have scaled to production. Restated in every 2026 analysis without new data.

Meanwhile, the bullish case has been updated three times in May alone — by the three companies with the best demand visibility on the planet.

Monday, May 4 — Amazon Q1 earnings. AWS discloses that its custom-chips business has crossed a $20 billion annual revenue run rate, and that Trainium revenue commitments now total more than $225 billion. Management linked the silicon strategy directly to agent inference, framing Trainium and Graviton as the substrate for the next generation of enterprise agent workloads. Three days later (May 7), AWS launches Bedrock AgentCore Payments in preview — the first managed payment capabilities purpose-built for autonomous agents, built with Stripe and Coinbase. Agents on Bedrock can now authenticate wallets, transact for APIs and MCP servers, and enforce spending governance — autonomously, mid-task.

Tuesday, May 19 — Google I/O. Sundar Pichai unveils Antigravity 2.0 as the centerpiece of the keynote — Google's agent-development platform, now a standalone desktop app for orchestrating cohorts of autonomous coding agents. In the headline demo, Antigravity builds a working operating system in 12 hours using 93 sub-agents, 15,000 model requests, and 2.6 billion tokens, for under $1,000 in API credits.

Wednesday, May 20 — Nvidia's Vera reveal. Jensen Huang tells investors he's identified a brand-new $200 billion market for Nvidia in agent compute. Nvidia has already booked $20B of Vera revenue in 2026. The pitch is explicit and workload-specific: as agents proliferate, GPUs alone don't serve the loop. Every agent needs a CPU socket next to its GPU for the orchestration phase — and Nvidia intends to own both. The Vera reveal also retroactively explains the $20B Groq acquisition from December 2025: fast inference is the rate-limiter on how quickly an agent loop can iterate. Nvidia bought the latency play because the latency constraint is becoming the binding one.

So which is it?

What are Nvidia, Google, and Amazon seeing in their order books that warrants moves of this size, and why is the analyst consensus on enterprise agent adoption so much more cautious than the actions of the people who actually sell the compute?

Both are true at once.

The pilots that fail are governance, integration, and ROI-measurement failures inside enterprises trying to deploy agents into messy organizational reality.

The compute being consumed is consumed by the agents that do work — coding agents inside Google, Anthropic's customers in production, AWS Bedrock's 125,000+ customers running Trainium-powered agents.

The aggregate demand curve goes up even when most projects fail, because the projects that succeed consume orders of magnitude more compute than chatbot deployments ever did. A 25% production success rate at agent token-burn rates is still a vertical demand curve.

One genuine uncertainty: consumer agents

The demand above is enterprise and developer agents. Klarna replaced 853 FTEs. JPMorgan runs 450+ production agents daily. Anthropic's 80x Q1 growth is enterprise-driven. Google's Antigravity 6x in eight weeks is Google's own engineers.

The consumer agent market is unproven. Google's Spark launch at I/O — a $100/month paywalled personal agent demoed with a block-party-planning use case — is exactly the kind of product that ships before its story is figured out. TechCrunch's Sarah Perez called it directly: Google launched the products before knowing how to sell them.

Private companies positioned for the agentic inflection

So apart from the hyperscalers, which private companies are positioned to benefit from the inflection?

Compute backbone — silicon and interconnect

Ayar Labs — silicon photonics / optical chip-to-chip interconnect. Agent workloads require many chips coordinating at low latency; copper-based interconnects bottleneck this. Intel, Nvidia, HPE among investors.

AI cloud & inference infrastructure for agent workloads

Lambda — Nvidia-aligned neocloud.

Crusoe — energy-flexible AI cloud, building gigawatt-scale data centers.

Modal Labs — serverless GPU + container platform with agent sandboxes.

Fireworks AI — fast inference platform. Inference latency is the rate-limiter on agent loops.

Baseten — model deployment. Production-inference layer for enterprise agents.

Agent orchestration & frameworks

LangChain — LangChain + LangGraph + LangSmith covers framework, orchestration, and observability. Broadest infrastructure footprint among pure-play agent companies.

Transaction & verification rails for agents

Stripe — Stripe Agent Toolkit is the most-used way for agents to charge cards.

Plaid — bank account connectivity. Every personal-finance agent needs Plaid.

Vertical and applied agents

Sierra — Bret Taylor's customer-service agent platform. Most enterprise-ready agent company today.

Glean — enterprise search and knowledge agents.

Perplexity — answer engine moving into agentic browsing and shopping. Consumer-facing wedge against Google Search.

Secondary market activity: the cohort is not a single market

The 12 companies above aren't trading like a single category. Liquidity, spreads, and premiums vary enormously, and the dispersion itself is informative.

Tightest two-sided markets: Crusoe (4% spread), Perplexity (7%), Ayar Labs (11%) — these are the names where institutional buyers and sellers are actually meeting at price.

Trading at premium: Lambda (+44.5%) and Crusoe (+49.5%) are well above their last primary rounds. Both are expected 2026 IPO candidates — the premium reflects pre-IPO accumulation, not just AI-infrastructure demand.

Trading at discount: Fireworks AI, Plaid (-52%), LangChain (-31%) — markdowns that range from "competitive pressure" (Fireworks against Baseten and the hyperscalers' inference services) to "model labs moving up the stack" (LangChain vs OpenAI Agents SDK and Anthropic's Claude Managed Agents).

Bids only, no asks at Sierra, LangChain, Fireworks, Modal — buyers want exposure; existing holders aren't selling.