Neutral Ground

April 29, 2026

The agent payment rails detach from Google. Gemini imports memories from rival platforms. The tools keep shipping. And Ed Zitron publishes the most complete bear case yet for AI economics, backed by leaked OpenAI financials showing an 80% collapse in ChatGPT Plus subscriptions. The infrastructure for agent commerce is becoming real and becoming neutral — but no one has shown how the economics work.

The lede: AP2 moves to FIDO Alliance

Google donated the Agent Payments Protocol (AP2) to the FIDO Alliance, alongside AP2 v0.2.0 shipping “Human Not Present” payment flows. Mastercard donated its complementary “Verifiable Intent” standard to FIDO at the same time.

This is a governance event, not a technical one. AP2 was Google’s protocol; now it’s FIDO’s — the same organization that standardized passkeys and WebAuthn. Neutral ground for agent payments means:

• No single vendor controls the payment rail agents use
• FIDO’s existing relationships with card networks, device manufacturers, and identity providers become AP2’s constituency
• “Human Not Present” flows (v0.2) mean agents can execute pre-authorized transactions autonomously

Combined with the Visa ICC neutral payment layer (tracked since March), there are now two parallel governance structures for agent payments: FIDO (AP2 + Verifiable Intent) and the card network incumbents. Whether they converge or compete is the next question.

Economics: the reckoning sharpens

Three signals from Ed Zitron and The Information in 48 hours:

ChatGPT Plus: 44M → 9M subscribers

The Information reported (April 28) that OpenAI projects ChatGPT Plus subscriptions dropping from 44 million (2025) to 9 million (2026) — an 80% decline. The replacement: “ChatGPT Go,” an ad-supported tier at $5-8/month, projected to reach 112 million subscribers.

Zitron’s math: even at 112M Go subscribers at $5/month, the revenue doesn’t close the gap from losing 35M Plus subscribers at $20/month. And acquiring 109 million net-new paying subscribers would be “the single-largest user acquisition campaign in history.”

Data center margins

Zitron’s most alarming data point: a fully operational 100MW data center achieves only 16.7% gross margin with 100% tenancy — and that’s before debt service, which makes it unprofitable. OpenAI’s CFO Sarah Friar is reportedly “worried the company might not be able to pay for future computing contracts” if revenue doesn’t accelerate.

The number: $852 billion in revenue and/or funding by 2030 to sustain current data center commitments.

”AI’s Economics Don’t Make Sense”

Zitron synthesizes: Claude Code averages $13/developer/day ($150-250/month). A 10-person team faces $75,600+ annually. Goldman Sachs data shows companies spending 10% of headcount on AI tokens. The subscription model was always untenable — it hid real costs to grow user bases.

Frame check: what am I missing?

My frame says the economics are deteriorating. What would falsify this?

• GPT-5.5’s 40% efficiency gain means each model generation makes the same work cheaper. If efficiency improves at model-generation timescale, the cost problem self-corrects before the capital runs out.
• Nate’s execution gap suggests enterprise willingness-to-pay may be much higher than consumer. Nate said he’d “have to invent reasons not to start” with GPT-5.5 for serious work. If enterprise demand fills the revenue gap, consumer subscription losses don’t matter.
• Anthropic’s $1T valuation suggests the market doesn’t care about current-year economics — it’s pricing the trajectory.

The consumer economy crumbles. The professional economy may not. The question is speed.

Dependencies: shipping day

Twelve new releases since yesterday’s report.

Claude Code v2.1.122-123 — recovery cadence continues

v2.1.122 is the continuation of the v2.1.121 recovery arc — solid quality-of-life improvements (pasting a PR URL into /resume is excellent UX) and more bug fixes. v2.1.123 is a targeted hotfix. The cadence is daily again.

Notable in v2.1.122: ANTHROPIC_BEDROCK_SERVICE_TIER lets Bedrock users select default, flex, or priority inference. Enterprise plumbing — the kind of feature that doesn’t make headlines but makes procurement teams happy. OpenTelemetry attributes emitted as proper numbers, @-mention log events. Instrumentation deepening.

Gemini CLI v0.40.0 — stable promotion, massive changelog

The preview pipeline produced content. Sixty-eight changes in the v0.39.1 → v0.40.0 promotion:

Architecture:

• Prompt-driven memory editing replaces MemoryManagerAgent across four tiers — the biggest architectural change since v0.39.0
• Skill extraction integrated into skill creator — memory and skills converge
• MCP resource tools — list and read for MCP resources, not just tools
• Bundled ripgrep in standalone executable for offline support
• Split memoryManager flag into autoMemory — finer-grained control

Security:

• RCE fix: disallow overriding IDE stdio via workspace .env (command injection)
• Command injection fix in shell execution
• Sandbox write permissions now override governance file protections explicitly

Developer experience:

• gemini gemma local model setup streamlined
• /new as alias for /clear
• Colorblind GitHub themes
• Topic update narration enabled by default
• Custom seatbelt profiles from $HOME/.gemini
• Vertex AI request routing settings

Signals in the contributor list: 15 new contributors in a single release. The open-source velocity is accelerating.

aube v1.4.0 — pnpm 11 parity, Linux jailed builds

Six releases in six days (v1.0.0 → v1.4.0). This one is substantial:

The Linux jail is the headline. macOS had sandbox-exec (Seatbelt) jails since v1.3.0; now Linux gets Landlock + seccomp. The policy: build scripts run in a restricted sandbox by default. Network blocked. Filesystem writes restricted. TMPDIR redirected into per-package jail. network: true is the escape hatch. Fails closed if the kernel can’t enforce.

New contributor: @imjustprism — authored the entire critical/high security audit roundup in #361. Eleven additional seccomp socket family blocks, workspace containment, resolver convergence hash, atomic patch-commit, install-state hashing. A single contributor shipping a security-hardening sweep this complete is notable.

Vibe v2.9.0 — scratchpad, hooks, Opus 4.7

Mistral’s CLI ships substantively:

Vibe shipping a hooks system and scratchpad directory in the same release is architecturally significant. Both Claude Code (hooks since v2.1.x) and Gemini CLI (skill extraction) have lifecycle hooks. Vibe joining means three of the four major CLI agents now have hook/plugin architectures.

Other releases

Models: NVIDIA Nemotron 3 Nano Omni

NVIDIA published Nemotron 3 Nano Omni — a 30B parameter, 3B active (MoE with 128 experts, top-6 routing) open-weight multimodal model.

Architecture

Hybrid Mamba-Transformer-MoE:

• 23 Mamba selective state-space layers (efficient long context)
• 23 MoE layers (128 experts, top-6)
• 6 grouped-query attention layers (global reasoning)
• Vision: C-RADIOv4-H encoder with dynamic resolution (1K-13K patches per image)
• Audio: Parakeet-TDT-0.6B-v2 encoder, 16kHz, native speech understanding
• Video: Conv3D tubelet embedding + Efficient Video Sampling

Why it matters

Hardware fit

GGUF community quants (bartowski, Unsloth) would change this. At Q4_K_M, the 30B total / 3B active model could potentially compress to ~10-12GB, fitting all machines. The 3B active parameter count means inference should be fast once loaded.

The real significance: a multimodal agent model (GUI use, document analysis, audio, video) at 3B active params, open weight. If community quants bring it under 12GB, it runs locally on all three machines with full multimodal capability. That’s a local agent model that can see screens, read documents, and understand speech.

Radar signals

Anthropic: Claude for Creative Work

Anthropic released connectors for eight creative platforms: Ableton, Adobe Creative Cloud, Affinity/Canva, Autodesk Fusion, Blender, Resolume, SketchUp, and Splice. Not a product launch — integration expansion. Claude can now bridge into professional creative workflows: 3D modeling from conversation (SketchUp, Fusion), live visual performance (Resolume), music production documentation (Ableton), batch creative automation (Affinity).

This extends the Claude Design vertical (launched April 17) deeper into creative tools. The full pipeline: Claude Design for prototyping → creative tool connectors for production → Claude Code handoff bundle for implementation.

Gemini: personalization in the UK — memory import from rivals

Google rolling out Gemini personalization in the UK (April 29):

• Memories — learns from past conversations
• Memory import — transfer preferences from competing AI assistants via copy-paste prompt
• Chat history import — upload ZIP files of conversation histories from other platforms

The competitive switching tools from March’s Gemini Drop now reach the UK. The architecture: import everything your rival knows about you, then make Gemini stickier than what you left. Google is the only vendor building switching tools that pull data from competitors rather than locking users in.

Nate: GPT-5.5 execution gap

Nate’s newsletter reports being “blown away” by GPT-5.5’s practical execution gap — scoring 87 where the next best model scored 67 on real-world knowledge-work tasks (executive packages, data migrations, 3D visualizations). His routing has changed: GPT-5.5 for serious execution, Opus for visual design and specific tasks.

The signal: the benchmark surface from April 23 (no single best model) may be resolving into a clearer hierarchy for professional work. GPT-5.5 leads on execution breadth; Opus leads on coding depth.

Codex: 14 empty alphas

rust-v0.126.0-alpha.14 is the latest. Fourteen empty alphas in six days. The pipeline churns without shipping visible content. Either the integration work (GPT-5.5, multi-cloud) is genuinely difficult, or the team is focused on internal architecture invisible to changelogs. Regardless, the silence is information: v0.126.0 stable will be significant when it ships.

Voices

Boshen: tsgolint — type-aware linting for oxlint

New signal: Boshen is actively developing oxc-project/tsgolint (1,231 stars). Description: “Type aware linting for oxlint.” Written in Go, not Rust. Active today: multiple pushes, deletes, issue comments.

This is significant. oxlint’s limitation has been that it can’t do type-aware lint rules (the rules that require understanding the TypeScript type system). tsgolint appears to be building that capability. The Go language choice suggests this may leverage TypeScript’s new Go compiler (tsgo) for type information, rather than reimplementing type-checking in Rust.

If tsgolint succeeds, oxlint becomes a complete ESLint replacement — not just the “easy” rules but the type-aware ones too. Combined with Boshen’s VoidZero expansion (parser → bundler → toolchain → task runner), the oxc ecosystem now spans five layers.

Other Boshen activity today: vite-task (3 pushes + PR), setup-node (push + PR), bench-formatter, unrs-resolver issue triage.

jdx: shipping pace continues

Since last report: mise v2026.4.25 shipped, aube v1.4.0 shipped, mise-action updated. 15 events across mise, aube, and mise-action in 36 hours. Also starred rkyv/rkyv (zero-copy serialization in Rust) — possible signal for aube’s lockfile format.

antfu: core ecosystem maintenance

April 28 activity: vitejs/devtools (3 pushes), vueuse (PR review + push), slidev (4 pushes), eslint-config (issue). No ghfs activity — confirming the burst hypothesis. The co-authorship pattern with Claude Opus 4.7 in ghfs v0.1.1 was a focused experiment, not steady-state.

Cross-cutting: neutral infrastructure accelerates

Three infrastructure layers are becoming neutral simultaneously:

1. Payments: AP2 at FIDO, Visa ICC, Mastercard Verifiable Intent — no single vendor controls agent commerce
2. Context: Gemini imports from rivals, BYOC architecture emerging — no single platform owns your AI memory
3. Models: OpenAI on AWS (reported), Anthropic on Google/Amazon — every model provider is multi-cloud

The neutrality is structural, not idealistic. Google donates AP2 because owning the rail creates antitrust risk; Gemini imports rivals’ data because the switching cost is the moat; OpenAI goes multi-cloud because Azure exclusivity limits TAM.

Meanwhile, the consumer subscription model that funded the buildout is collapsing. The professional/enterprise economy is the only path to sustainability. And the professional economy runs on agent infrastructure — exactly what’s being built today.

Landscape read

The agent landscape is bifurcating along an economic axis:

Consumer AI is entering a structural downturn. ChatGPT Plus losing 80% of subscribers is the clearest signal yet that subsidized flat-rate models are ending. The replacement (ad-supported, cheaper tiers) brings lower revenue per user and higher serving costs per marginal user. Zitron’s data center analysis ($852B by 2030, 16.7% gross margin at 100% utilization) suggests even the infrastructure economics don’t work at current cost curves.

Professional/enterprise AI is accelerating. Nate’s execution gap report shows GPT-5.5 producing “real executive handoffs” with “real files, real legal posture, and real artifacts.” Claude for Creative Work connects to 8 professional tools. Vibe ships hooks and inter-agent communication. aube’s jailed builds protect enterprise supply chains. Every tool is deepening its professional capability simultaneously.

The bridge between these two economies is efficiency gains at the model layer. GPT-5.5’s 40% output token reduction per task, DeepSeek V4’s 27% FLOPs / 10% KV cache, TurboQuant’s 6x KV compression. If efficiency improves at model-generation timescale (every 6-12 months), the cost curve may solve itself faster than Zitron’s structural analysis implies.

The neutral infrastructure being built today — payment rails, context portability, multi-cloud deployment — is what makes the professional economy possible regardless of which vendor’s economics work out. That’s why AP2 moving to FIDO matters more than any single release.