AI & Tech Brief ⚡

📌 Navigate

📊 Exec Summary

Anthropic shipped a confirmed product triple and had a model and IPO timeline leak in the same week — a combination of verified engineering releases and press-reported speculation that should be read separately.

Five things moved in AI/tech this week:

1. Anthropic ships auto mode, Cowork computer use, and Dispatch in a single week — model-based permission classifiers replace human approval loops in Claude Code
2. Claude Mythos revealed via accidental data cache — "step change" in capabilities per leaked draft, cyber-offense concerns; no official announcement; benchmark data absent
3. Anthropic targets October IPO at $400-500B — The Information reporting, not confirmed by Anthropic; $19B annualized revenue figure also press-reported, not verified in IR filings
4. H100 rental prices reverse course and melt UP — analyst and podcast observations, not exchange-traded price data; directionally consistent across sources
5. Stripe launches Projects.dev — CLI-first agent provisioning announced by Patrick Collison on X with live product

The pattern: Agentic infrastructure is splitting into two races — who controls the permission layer (Anthropic's classifier vs. OpenAI's sandbox) and who controls the provisioning layer (Stripe's CLI vs. cloud-native toolchains). GPU scarcity is the constraint binding both. The IPO and Mythos signals are speculative overlays on top of confirmed product moves — read them as such.

1️⃣ Anthropic ships auto mode, Cowork computer use, and Dispatch in a single week

TL;DR: Anthropic released three tightly coupled agentic features — Claude Code auto mode (model-based permission gating), Cowork computer use (desktop control for non-coding tasks), and Dispatch (phone-to-desktop session triggering) — collapsing the distance between "copilot" and "autonomous agent."

What happened

• Claude Code auto mode shipped March 24. It replaces binary permission prompts with a Sonnet 4.6-based transcript classifier that evaluates each proposed action against decision criteria before execution.
• Two classifier layers: a read classifier (what Claude can see) and a write classifier (what Claude can do). Prompt injection defense is built into the write layer.
• Cowork computer use rolled out March 25. When no API connector exists for a service (e.g., a proprietary internal tool), Claude falls back to direct mouse and keyboard control. Google Workspace and Slack connectors are prioritized.
• Dispatch shipped alongside Cowork. Users type a prompt on their phone; it triggers a full Claude Code or Cowork session on their desktop — same folders, same context.
• Zvi Mowshowitz called it "whatever else you think about Anthropic's agentic coding department, they ship."

📊 Technical specifics

🔗 Primary source → Fortune exclusive

🔍 The non-obvious point

The leak itself may matter more than the model. Anthropic's entire brand is built on responsible scaling — and an accidental public data cache is a basic operational failure, not a frontier safety problem.

• The defender-first rollout framing is strategically useful regardless of whether Mythos is actually that dangerous. It positions Anthropic as the only lab that voluntarily gates its own product for safety reasons — exactly the narrative needed heading into an IPO and a regulatory environment. Whether the framing reflects genuine capability is unknown.
• The cyber-offense framing also creates a moat narrative. If Mythos is genuinely ahead in vulnerability detection, enterprise security teams have a reason to choose Anthropic. But the claim rests entirely on leaked draft copy with no accompanying evals.
• The absence of benchmark data is notable and distinguishes this from every prior Claude release. This item should be held as speculation until official documentation is published.

⚠️ What would confirm this item: An official Anthropic system card, published evals, or formal product announcement. Until then, treat capability claims as unconfirmed.

👀 What to watch

• Whether Anthropic formally announces Mythos before the S-1 filing (expected late summer). A "most powerful model" claim is material for IPO valuation.
• Whether competing labs (OpenAI, Google) respond with their own cyber-capability claims or dismissals.

3️⃣ Anthropic targets October 2026 IPO at $400-500B valuation

TL;DR: The Information reported on March 27 that Anthropic is targeting an October 2026 IPO, with Goldman Sachs and JPMorgan Chase engaged as lead banks. Revenue and valuation figures in this item are press-reported estimates — not confirmed in audited financials or SEC filings.

What happened

• The Information reported on March 27 that Anthropic is targeting an October 2026 IPO window. Anthropic has not confirmed this publicly.
• Goldman Sachs and JPMorgan Chase are reported as engaged as lead banks; Wilson Sonsini as legal counsel. Press-reported, not confirmed by any party.
• S-1 filing expected late summer 2026; roadshow and Nasdaq listing targeting Q4. Press-reported timeline.
• Annualized revenue of $19B by March 2026 (up from $9B at year-end 2025 and $1B in December 2024) is a press-reported figure, not a disclosed financial result. The growth rate is extraordinary and warrants independent verification when the S-1 is filed.
• Last private valuation of $380B (Series G, February 2026) and reported private offers at $800B are secondary-market signals, not confirmed transaction prices.

📊 Reported metrics

🔗 Primary source → The Tech Portal Also: TradingKey analysis, TechCrunch on investor dynamics

🔍 The non-obvious point

The revenue trajectory — $1B to $19B annualized in 15 months, per press reports — would be the single most important number in AI right now if confirmed in a filed document.

• This growth rate, if verified, would be faster than any enterprise software company in history at comparable stages. The S-1 is the verification event. Until then, these figures should be read as analyst projections consistent with observed product growth, not settled fact.
• The $800B private offers (more than double the Series G) suggest some investors believe Anthropic could be worth more than OpenAI at IPO. TechCrunch reports that "Anthropic's rise is giving some OpenAI investors second thoughts."
• The Mythos leak timing is suspicious. An accidental leak of a "step change" model one day before IPO banker engagement leaks is either genuinely accidental or a masterclass in narrative sequencing. Both items are press-reported and neither is confirmed.

👀 What to watch

• S-1 filing: the first confirmed data point in this chain. Revenue and valuation figures become verifiable at that point.
• Whether OpenAI accelerates its own IPO timeline in response.

4️⃣ H100 rental prices reverse course and melt UP

TL;DR: After a two-year depreciation cycle that bottomed out during the DeepSeek R1 shock, H100 rental prices have risen sharply since December 2025, driven by inference demand that now outstrips supply. This is a directional trend signal from analyst commentary, not a reported price dataset.

What happened

• Latent Space (Swyx) reported on March 28 that H100 rental prices have gone "VERY up" since December 2025. This is Swyx's observation, not a published price index.
• The original depreciation cycle (2022-2025) was driven by post-hype demand normalization and the DeepSeek R1 shock.
• The reversal is driven by inference scaling: production AI workloads now consume more GPU-hours than training. This is a consistent analyst thesis, not a measured dataset.
• Dylan Patel on Dwarkesh Podcast stated H100s are "worth more now than at launch." Podcast commentary — not a structured price survey.
• Matthew Sigel's chart shows rental rates climbing above $3/GPU-hour, approaching 2023 peaks. Secondary analyst data point.

🔗 Primary source → Latent Space: H100 prices are melting UP

🔍 The non-obvious point

GPU price inversion changes the economics of every AI startup's build-vs-buy decision.

• If inference costs are rising rather than falling, the "just use the API" strategy becomes more expensive over time, not less. This favors companies that own or have long-term contracts on GPU capacity — and penalizes API-dependent startups.
• NVIDIA's Vera Rubin platform (announced at GTC, shipping H2 2026) promises 10x inference cost reduction vs. Blackwell. That is a confirmed announcement, but the timeline is 6+ months away. The interim scarcity creates a window where GPU-rich companies have a structural advantage.
• For biotech AI specifically: inference-heavy workloads like molecular dynamics simulations, protein folding ensembles, and large-scale virtual screening are now more expensive to run on spot-market rentals.

👀 What to watch

• NVIDIA Q1 FY2027 earnings (late May) for datacenter revenue and order backlog — the first reported data that can confirm or deny the rental-price trend.
• Vera Rubin production timeline — any delay extends the scarcity window.

5️⃣ Stripe launches Projects.dev — CLI-first agent provisioning

TL;DR: Stripe shipped Projects.dev on March 27, enabling agents to provision third-party services via CLI commands (e.g., stripe projects add posthog/analytics), positioning the terminal as the universal agent interface.

What happened

• Patrick Collison announced Projects.dev on X on March 27.
• The product lets agents run a single CLI command to create accounts, obtain API keys, and configure billing for partner services.
• Launch partners include PostHog, Vercel, Supabase, and several other developer infrastructure companies.
• Stripe handles the billing relationship, making it the payment rail for agent-provisioned services.
• Latent Space covered it as the headline story for the week, framing it as "Everything is CLI."

📎 Sources

Sources of truth

Also consider reading