In Issue #009, we wrote: “Watch what Anthropic does with compute allocation over the next 60 days.” It took 14.
On Thursday, Anthropic published a post-mortem confirming what users — and Stella Laurenzo’s 6,852-session AMD telemetry — had been saying for over a month: Claude Code degraded between March 4 and April 20. Three changes shipped in that window. Only one was a classic bug. The other two were product-layer tradeoffs — reduced reasoning effort and reduced verbosity — that saved latency or usage but degraded the developer experience badly enough that Anthropic reversed them. Then Anthropic locked in 10 gigawatts from Amazon and Google in five days, OpenAI shipped a new model nobody asked for at twice the price, and 92,000 tech workers got the news that AI is now writing 65% of the code at companies that just laid them off. Speed and fear became the same vector.
🧠 Anthropic Ate Some Humble Pie
The compute crunch isn’t theoretical — it’s already shipping in your IDE
The post-mortem named three causes. Mar 4: default reasoning effort dropped from “high” to “medium” — Anthropic later called the tradeoff “bad” themselves. Reverted Apr 7. Apr 16: system prompt instructions capped output to 25 words between tool calls and 100 words for final responses — their own broader evals showed a 3% drop in coding quality. Reverted Apr 20. Mar 26: a caching bug cleared thinking blocks every turn instead of once. Users reported “forgetfulness, repetition, odd tool choices.” Cache misses compounded into faster usage-limit drains. Fixed Apr 10.
Two of the three weren’t classic bugs. They were product-layer tradeoffs that saved compute and hurt quality, defended publicly until the receipts came in. Anthropic also briefly pulled Claude Code from the $20 Pro plan on April 21 — described as a “pricing test on 2% of new signups” — and shifted enterprise pricing toward consumption-based, which one analyst estimated could triple costs for heavy users.
Two weeks earlier, Bloomberg and CNBC reported that OpenAI sent a memo to its own investors calling Anthropic “compute constrained” and projecting 30 GW by 2030 against Anthropic’s 7-8 GW by end of 2027. Four days after that, OpenAI’s CRO Denise Dresser sent a separate memo to staff (viewed by CNBC and The Verge) accusing Anthropic of inflating revenue by $8 billion through gross cloud accounting. Anthropic’s response shipped 11 days later.
Why it matters: When labs are spending tens of billions on raw silicon, the only short-term margin knob left is per-request compute allocation. The moment it gets squeezed, your developer experience gets thinner. We told you. The receipts came in.
Hype vs. Reality: 9/10. The story isn’t that Claude Code got worse. It’s that the AI lab most associated with safety transparency spent a month publicly defending changes their own evals later flagged as quality drops. Trust is the asset. This week it took a hit.
⚡ Three Deals, Five Days, Ten Gigawatts
Amazon, SpaceX, and Google all wrote checks to the agentic stack inside one work-week
April 20 — Amazon → Anthropic. Per Reuters, $5B now plus up to $20B more, plus a $100B AWS commitment over 10 years for 5 GW of Trainium2/3/4 capacity. Anthropic now runs on 1M+ Trainium2 chips. April 21 — SpaceX → Cursor. Reuters reported a right to acquire Cursor for $60B later this year, or pay a $10B “collaboration fee.” Cursor halted a $2B fundraise to take it. April 24 — Google → Anthropic. TechCrunch and Bloomberg reported $10B now at a $350B valuation, plus up to $30B more on milestones, plus another 5 GW of TPU capacity starting 2027.
Anthropic is now the only frontier model deployed across AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry — with new gigawatt-scale compute commitments from Amazon and Google. ARR run-rate is $30B, per Bloomberg reporting, up from $9B at end of 2025, $14B in February, $19B in March. Secondary markets are pricing the company at $1T. The Information reports the IPO target is October at $400-500B valuation, $60B+ raise. Eight of the Fortune 10 are Claude customers.
This is what an answer to OpenAI’s compute memo looks like. OpenAI argued “we have more power plants.” Anthropic’s response is “we’ll build them too.”
Why it matters: GPU pricing volatility is no longer a 2027 problem. If you’re building exclusively on one frontier provider, the compute crunch is now your problem too. Audit your stack lock-in this week, not next quarter.
Hype vs. Reality: 8/10. The dollar amounts are real. The 10 GW are real. What’s not yet real is the gap between commitments and delivered capacity — most of this lights up in 2027.
🤖 Enterprise Agent Day Was Real
April 22 was the day the shipping spree from Issue #010 became GTM
In a single 24-hour window: OpenAI launched Workspace Agents in ChatGPT — Codex-powered, Slack-integrated, with approvals, Compliance API visibility, and prompt-injection safeguards, free until May 6. Google rebranded Vertex AI as Gemini Enterprise and split the TPU line for the first time — TPU 8t (Sunfish, Broadcom, training) and TPU 8i (Zebrafish, MediaTek, inference), 9,600-chip superpods, 2.7x perf-per-dollar. Salesforce-Google Agentforce-Gemini announced the same afternoon.
Then on April 23, GPT-5.5 (“Spud”) dropped — six weeks after 5.4. 82.7% Terminal-Bench 2.0, 78.7% OSWorld-Verified, but the API price doubled to $5/$30 per million tokens. The builder verdict landed fast. Developer Theo Browne’s widely-shared review summed it up: “the smartest model you’ll hate working with.” OpenAI’s own developer guidance for GPT-5.5 told users to start prompts from scratch — don’t carry over 5.4 prompts. Role definitions are back at the top of the system prompt.
April 24 — DeepSeek V4 dropped. V4-Pro is 1.6T parameters (49B active), MIT-licensed, 1M context native, with 9.5x lower KV-cache memory than V3.2. $1.74/$3.48 per million tokens — 7x cheaper than Claude Opus at near-identical SWE-bench (80.6% vs 80.8%). V4-Flash at $0.14/$0.28 is cheaper than GPT-5.4 Nano. Trained on Huawei Ascend. Same day, Reuters broke that the State Department issued a cable warning diplomatic posts about Chinese AI distillation. SMIC stock jumped 10%. Simon Willison’s verdict: V4-Pro is “the cheapest of the larger frontier models.”
Two named devs, opposite reactions, same week. GPT-5.5 wins benchmarks but loses on UX. DeepSeek V4 ships at one-seventh the price with weights you can run yourself.
Why it matters: The China stack is no longer “the cheap option.” It’s the post-Nvidia option. If cost is a real constraint in your stack, pull DeepSeek V4 this week and run it against your top three workflows.
Hype vs. Reality: 7/10 on GPT-5.5. Wins benchmarks, costs more, requires prompt rewrites. 9/10 on DeepSeek V4. Math + geopolitics both real.
🔒 Mythos Collapsed Three Ways in 14 Days
The “we kept it tight” thesis from Issue #009 is now formally dead
The NSA is using Mythos Preview (Axios scoop, April 19-20) — despite the Pentagon, its parent agency, having Anthropic on a supply-chain risk blacklist. The Pentagon is simultaneously arguing in federal court that Anthropic’s tools threaten national security. UK intelligence agencies also have access via the country’s AI Security Institute. Amodei met White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent on Friday April 17. Both sides described it as “productive.”
A Discord group breached Mythos on launch day through a third-party vendor by guessing the API URL based on Anthropic’s naming conventions. They’ve used it continuously since, per multiple reports. Anthropic restricted Mythos to roughly 40 organizations and only publicly named 12. Gating the most dangerous AI model in commercial history failed through predictable access patterns and a third-party vendor environment.
DHS demoed jailbroken AI to all House lawmakers on April 22. The closed-door briefing introduced new vocabulary: “abliterated” (refusal mechanism deactivated) versus “censored” (Claude, ChatGPT). Members got live demonstrations of nuclear bomb instructions and attack plans. Rep. Gabe Evans (R-Colo.) told Politico: “models without safeguards ‘gave answers to all of those things.’”
Same week, Florida AG James Uthmeier expanded the state’s OpenAI probe to a criminal investigation (April 21). The FSU shooter (April 2025) consulted ChatGPT about gun selection, ammunition, and timing before killing two people. Uthmeier: “if it was a person on the other end of that screen, we would be charging them with murder.” Subpoenas served. Per court records, trial begins October.
Why it matters: The story is no longer whether American labs can be safety stewards in theory. It’s whether they can stay credible as the regulatory floor moves under them.
Hype vs. Reality: 10/10. Containment was the entire premise of Glasswing. It’s gone.
🚨 Workers Are Training Their Replacements
92,000 tech layoffs in the first four months of 2026 — and counting
The April 23 doubleheader landed harder than anything that came before it. Meta announced 8,000 layoffs (10% of workforce) plus 6,000 canceled roles, effective May 20, to fund $115-135B in 2026 capex (up from $72.2B in 2025). Per The Next Web’s reporting, hours before the layoff memo leaked, Meta awarded six top executives stock options worth up to $921M each, tied to a $9T market cap target by 2031. The same day, Microsoft opened voluntary buyouts to ~7% of US workforce (~8,500 employees) — first time in 51 years. Eligibility: age + tenure ≥ 70. Microsoft stock dropped 4% on the news, per CNN.
The pattern from earlier in the month was already clear. April 15: Snap fired 1,000 employees (16% of workforce). CEO Evan Spiegel cited “rapid advancements in AI.” AI now writes 65% of new code at Snap, per the company’s own SEC filing. CNBC reported the stock jumped roughly 8% the same day. April 21: Meta’s Model Capability Initiative rolled out keystroke, mouse, and screenshot surveillance for AI training across the workforce. April 24: GitHub Copilot training-by-default went live. Free, Pro, and Pro+ users opted in unless they manually opt out. Business and Enterprise exempt. Microsoft is now using individual developers’ code to train models that compete with their employers.
Anthropic’s own 81,000-user economic survey dropped April 23 with the cleanest U-curve in any AI labor study to date.
The workers reporting the largest AI-driven productivity gains expressed the highest displacement anxiety.
One in five voiced concern about economic displacement. Top 25% AI-exposure mentioned the worry 3x more often than the bottom 25%. Mean productivity rating: 5.1/7 — “substantially more productive.”
Speed and fear are the same vector now.
WEF data, cited by former UK Prime Minister Rishi Sunak this month: US entry-level job postings down 35% in 18 months.
Why it matters: The 81K survey just gave every CFO in America a reason to revisit headcount. Document your AI productivity gains before your CEO does. And opt out of GitHub Copilot training — github.com/settings/copilot/features — before your code becomes someone else’s competitive advantage.
Hype vs. Reality: 10/10. Anthropic published this. Their own users said it.
📡 Quick Signals
Cognition AI talking $25B — Bloomberg reported on April 23 that the AI coding firm is in early talks to more than double its valuation. Devin ARR went $1M (Sep ‘24) → $73M (Jun ‘25). Inbound surged after the SpaceX-Cursor deal. The market is now pricing agents that close tickets, not copilots that suggest.
Tesla Terafab + Intel 14A (4/22) — Tesla becomes Intel’s first major external 14A customer. Targeting 1 TW/year of compute (vs ~0.5 TW total US capacity now). Bernstein estimates $5-13 trillion capex required. Intel +3.6% after-hours.
A Karpathy-inspired CLAUDE.md skills repo crossed 70K stars — distilling Karpathy’s public LLM coding observations into reusable config guidance. Skills-as-knowledge-distribution is now a category.
Cursor + Chainguard partnership (4/21) — security infrastructure for AI-generated dependency decisions. “AI agents are picking packages at machine speed; humans cannot manually review.”
Defense AI is its own funding vertical now — Shield AI hit $12.7B valuation (+140% in a year). Anduril, Helsing, and Cognition recruiting from the same ML bench as OpenAI and Anthropic. HR analytics firms are flagging upward pressure on engineering compensation as defense-AI rounds compete for the same talent pool.
OpenAI ChatGPT for Clinicians (4/22) — free for verified US clinicians, HIPAA support. 72% of physicians now use AI in clinical practice (up from 48% in 12 months). No institutional contract required — bypassing CIOs entirely.
Anthropic-NEC Japan (4/24) — 30,000 employees deploying Claude. First Japan-based global partner. Geographic redundancy is now procurement strategy.
Maine’s governor vetoed a proposed pause on data centers over 20MW (Reuters, 4/24), but the bill’s progress shows the data-center backlash has moved from rhetoric into legislative machinery. Saline, Michigan’s Oracle/Stargate facility secured $16B in funding the same week. Rep. McIver introduced the AI Data Center Site Selection Transparency Act. The data center bills wave from #009 just got teeth.
Musk v. Altman trial opens Tuesday — jury selection Monday April 27 in Oakland federal court, opening arguments April 28. $134B at stake, narrowed to two surviving claims (breach of charitable trust + unjust enrichment). Witnesses: Musk, Altman, Brockman, Nadella, Murati, Sutskever. Liability phase runs through mid-May.
EU AI Act main enforcement opens August 2 — 97 days from Monday, April 27. Annex III high-risk regime covers HR/recruitment/workforce-monitoring AI, even as Omnibus-delay negotiations continue in Brussels. Map your exposure now.
🎯 The Playbook
Your move this week
- Audit your agent stack lock-in. If you’re building exclusively on one frontier provider, the compute crunch is now your problem too. Hedge.
- Document your AI productivity gains before your CEO does. The 81K survey just gave every CFO a reason to revisit headcount.
- Opt out of GitHub Copilot training. github.com/settings/copilot/features. Don’t feed the model that competes with you.
- Pull DeepSeek V4 if cost is the constraint. 7x cheaper than Claude Opus at near-parity SWE-bench. The geopolitics are real. So is the math.
- Map your EU AI Act exposure today. 97 days to August 2. HR, recruitment, and performance management AI are high-risk under Annex III. Quality management system, technical documentation, conformity records — all of it has to exist before that date.
🔥 What’s Viral Right Now
The Anthropic post-mortem screenshots — circulating on every AI Twitter account this weekend. Devs who got dismissed for a month are taking victory laps with Stella Laurenzo’s name in the captions.
The Karpathy-inspired CLAUDE.md repo at 70K stars — there’s a clear appetite for “how to actually configure these things.” Skills-as-knowledge-distribution is now a category, not a one-off.
The GitHub Copilot opt-out countdown clocks going up everywhere. Microsoft picked the worst possible week to flip that switch.
We told you to watch the compute. The receipts came in.
Stay building. 🛠️
— Matt