AI Waves #7. Stripe shipped the wallet, the NSA used the model, and Anthropic locked in $65B
Every contested layer of the AI stack got claimed by an incumbent
AI Waves #7 | May 1, 2026 | Nazaré Ventures
Previous issues: #1 | #2 | #3 | #4 | #5 | #6
Two weeks ago I argued in The New Users that agents are the customer software now needs to rebuild itself around, and that every layer of the stack agents need, from compute and memory through coordination, identity, payments, and tooling, has to be rebuilt from scratch. Most of the companies that will do that work, I wrote, are still being founded.
This week, two of those layers shipped at once. Stripe held its annual conference and launched the Machine Payments Protocol. Mastercard and Google quietly extended their parallel agent-commerce stack. The Anthropic story took its sharpest turn yet, with the NSA reportedly using a model the lab refused to ship publicly, the White House drafting a workaround to the Pentagon ban, and $65B in fresh hyperscaler compute commitments closing in four days at a reported $900B valuation. GPT-5.5 also shipped, treated below alongside the AGEI math.
Software begins to repackage
Stripe ran its annual Sessions conference on April 29 and 30, shipping 288 launches across the two days. The headline release is the Machine Payments Protocol, co-authored with Tempo (Stripe and Paradigm’s payments-focused L1, mainnet March 18). MPP is HTTP 402-based, agent-native by design, and rail-agnostic at the spec level: stablecoins settle on Tempo, fiat settles via Shared Payment Tokens against existing card and bank rails, and crypto deposits clear directly. The agent makes a request; the protocol negotiates the rail in milliseconds based on what the merchant accepts.
Streaming payments shipped on the same plumbing. Stripe’s own description: “stablecoin micropayments on the Tempo blockchain,” combining Metronome metering with Tempo settlement so that an LLM API gets paid the instant a token is generated rather than at the end of a billing month. This is the AI-native business model the industry has been describing for two years, and Stripe is shipping it on chain. Link’s agent wallet sits one layer up at the consumer surface and is fiat-first today, with one-time virtual cards issued per agent task and settled through cards and banks where 250M existing Link users already are. The release notes also state plainly that “support for agentic tokens, stablecoins, and other payment types are coming soon.”
The cleaner reading of Stripe’s week is one of absorption rather than displacement. Stripe absorbed the rails crypto bet on, kept its existing distribution, and shipped a multi-rail spine in which stablecoins handle the machine-to-machine streaming layer and fiat handles the consumer-facing wallet. The two-year crypto narrative was correct about the rails it would take, and incorrect about the company that would lay them down. Patrick Collison has framed stablecoins as “room-temperature superconductors for financial services,” and at Sessions he called AI “the biggest platform shift for the economy since the internet,” adding that “in the not-too-distant future agents will account for most transactions online.”
The other card network shipped a parallel agent-commerce stack with much less press. Mastercard’s Verifiable Intent, co-developed with Google, is a cryptographic proof framework that ties identity, intent, and action together for every agent transaction. It rides on top of Google’s AP2 and UCP, defining what an agent is authorised to do, what it can buy, and how its actions are cryptographically attested. On April 16, Crossmint integrated Mastercard Agent Pay and Verifiable Intent into Lobster.cash, putting that stack into production for the more than 1M agents on OpenClaw alone. The rails are fiat and Mastercard’s network, but AP2’s roadmap adds push payments and RTP, and Coinbase is in the partner list, so the crypto extension exists, it just lives one spec away.
The contested layer was never payments versus crypto. It was who could ship a multi-rail agent-native settlement spec with the merchant network already attached. Two card networks shipping agent-native primitives in the same month is the incumbent payment networks recognising that the rails work and racing each other to absorb them.
The Anthropic triangle
The week’s second arc is that the political, financial, and procurement layers around Anthropic all hardened in the same direction at the same time.
The Mythos thread is where to begin. Anthropic announced Mythos Preview on April 7 alongside Project Glasswing. Mythos is the lab’s frontier model above Opus 4.7, priced at $25 per million input tokens and $125 per million output tokens, roughly five times Opus 4.7. Anthropic explicitly will not release Mythos publicly. NBC News called the decision “the first time in nearly seven years that a leading AI company has so publicly withheld a model over safety concerns.” Initial access goes to eleven launch partners across the major hyperscalers, security vendors, and JPMorgan Chase, with $100M in usage credits and $4M to open-source security organisations on a 90-day disclosure window. Mythos has already found a 27-year-old OpenBSD flaw and a 16-year-old FFmpeg bug missed by automated tooling 5 million times.
Then on April 19, Axios reported that the NSA is using Mythos despite the Pentagon’s February supply-chain risk designation against Anthropic. Bloomberg followed on April 30 with the specifics: NSA officials are testing Mythos to find vulnerabilities in Microsoft products, and have been “impressed by its speed and efficiency in searching for potential security flaws.” The AGEI essay predicted governments would demand access to withheld frontier models for national security and eventually get it. Two weeks later, an intelligence agency is using a withheld frontier model in defiance of an active executive blacklist.
The political negotiation hardened around the same fact pattern. White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent met Dario Amodei in mid-April for what Axios called “a productive introductory meeting.” On April 29, Axios reported that the administration is drafting executive guidance to walk back the supply-chain designation and allow federal agencies to bring Mythos in. Trump told CNBC that Anthropic is “shaping up” and could “be of great use.” The Pentagon designation has not been lifted - DOD CTO Emil Michael said this morning that Anthropic remains flagged as a supply-chain risk - but Mythos itself is being routed around the designation as a separate national security matter.
Underneath the political traffic, Anthropic locked in $65B in compute commitments in four days. On April 20, Amazon committed $25B to Anthropic, with $5B invested immediately and up to $20B contingent on commercial milestones. In return, Anthropic committed more than $100B over the next ten years to AWS, securing up to 5GW of new capacity, and now runs on more than 1M Trainium2 chips. On April 24, Google committed up to $40B, with $10B in cash now and up to $30B contingent on performance targets, plus 5GW of GCP compute over five years. Both rounds priced off Anthropic’s February Series G ($350B pre-money, $380B post-money), so they are the same valuation reported under different conventions rather than a markdown between them. Anthropic’s run-rate revenue is now above $30B, up from $9B at the end of 2025, and CNBC reported on April 29 that the lab is in talks to raise at a $900B valuation, with the round expected to close in May. The fund mechanics are circular by design: cash flows from Amazon and Google into Anthropic and from Anthropic back into AWS and GCP, with milestone gating that keeps the equity release tied to compute deployment on each cloud.
The structure narrows the field. Anthropic is now anchored across three silicon platforms (Google TPU, Amazon Trainium, and Nvidia GPU) plus the December 2025 Hut 8 lease in Louisiana, and the next frontier lab with hyperscaler ambitions has to start either inside Microsoft, Meta, or Oracle, or from a non-American hyperscaler entirely. The public’s model is Opus 4.7, the procurement architecture is being rebuilt around Mythos, and the next $65B in compute is paying for whatever Anthropic ships after that.
GPT-5.5 ships
OpenAI shipped GPT-5.5 on April 23, priced at $5 per million input tokens and $30 per million output tokens, with a 1M-token context window. It is OpenAI’s first fully retrained foundation model since 4.5 and currently sits at #1 on the Artificial Analysis Intelligence Index at a score of 60. It is also expensive enough that sustained agentic use sits well above the cost ceiling for most workloads.
This is the AGEI math from last week’s essay landing on cue. The closed frontier still pushes ahead, and the lab releasing it captures a narrow band of the highest-stakes work where price is not the constraint. The middle tier (Opus 4.7 and a forthcoming GPT-5.5-derived public model) carries general consumer use. The open ecosystem (DeepSeek V4, Qwen, Kimi, and GLM) carries everything sensitive to either cost or sovereignty, which is most of what gets deployed in production. GPT-5.5 is the new ceiling. Most workloads will continue routing around it.
Quote of the Week
Andrej Karpathy, June 5, 2021
Five years later, Stripe and Mastercard are shipping the genuine innovation Karpathy named on rails Karpathy was sceptical the surrounding culture would let mature. Tempo is a stablecoin L1 inside a Stripe product. The packaging changed.
Portfolio
Dimensional. Stash Pomichter announced SpatialMemory2 on April 29 (, an open-source memory system for robots that builds a multimodal latent-space data store and lets agents run spatial and temporal queries against it asynchronously at runtime. Robots in production generate thousands of hours of video, lidar, and odometry per deployment. Stateless agents waste compute relearning what they already saw. Type “plant” against SpatialMemory2 and the system returns every moment the robot saw a plant, with map coordinates attached, alongside composable streams for velocity, brightness, and timing.
Prime Intellect. Prime Intellect launched FrontierSWE on the Environments Hub on April 29. FrontierSWE is a long-horizon coding benchmark by Proximal Labs. Agents run for an average of 11 hours per task and fail to solve almost all of them. Prime Intellect co-designed the granite_inf task, which measures an agent’s ability to optimise the speed of the forward pass of IBM’s Granite-Mamba-2 layer end to end. Only two runs reached near parity with the hidden human-engineered baseline: one Codex run and one Gemini run close behind. The Environments Hub now hosts more than 1,000 contributed environments. With Cursor reportedly in talks to raise $2B at a $50B+ valuation last week, the open infrastructure for training and evaluating long-horizon coding agents is the next contested layer. Prime Intellect already owns the open side of it.
Provably.Provably sits in the path of two threads in this issue. Mastercard’s Verifiable Intent above and the MCP RCE disclosure from last week both define the same problem: agents need cryptographic proof of what they did, what they queried, and which endpoints they touched. Provably’s qedb stack and agent guardrails SDK do precisely that, intercepting every API call, database query, and MCP server interaction and turning each into a verifiable proof downstream agents can check in milliseconds.
Intelligent Internet. Founder Emad Mostaque keynoted RenderCon 2026 on April 16 with a talk titled “How to Build the Holodeck”, breaking down how diffusion models are now learning real-world physics, why open-source world models are arriving faster than expected, and why compute access becomes the critical layer once these experiences scale. The line that travelled:
“I think that the holodeck is near again, minus the hardlight holograms. We’ll be able to have that, with our 3D glasses, with ocular occlusion and more.”
Open frontier capability plus open-source world models plus the compute substrate to run them defines a layer the hyperscalers do not yet own. Both worlds keep growing, with only one captured by existing distribution.
Closing
This week, the agent stack landed in public and the closed frontier hardened around Anthropic. The operators worth following are the ones still cataloguing which surfaces have been captured by whom and which haven’t. The buildout is not finished; the calendar of capture is just becoming legible.