Anthropic's Compute Miscalculation Breaks Its Flywheel
Anthropic's cautious capex stance left them compute-starved amid exploding agentic demand, triggering quota cuts, uptime woes, and confusing policies that drive users to OpenAI.
Anthropic's Elite Coding Flywheel—and Why It's Wobbling
Anthropic built the most efficient business flywheel in AI: superior coding models attract enterprise revenue, yielding proprietary coding data for training better models, funded by compute purchases that scale serving and training. This loop propelled Claude's Opus (4.6-4.7) to daily dominance for coders. But compute scarcity shattered it. Demand from agentic workflows—running 10 parallel agents nonstop—exploded beyond projections, especially via third-party harnesses like OpenClaw (best with Opus for tool calling and personality).
The core error traces to CEO Dario Amodei’s 18-month-old calculus: fearing OpenAI-style bankruptcy from overbuying data centers. He rejected trillions in projected capex, assuming 10x annual revenue growth might falter to 5x, dooming the firm if revenue hit $800B instead of $1T by 2027. > "If my revenue is not a trillion dollars... there's no force on earth that could stop me from going bankrupt if I buy that much compute."
Reality flipped the script: demand matched or exceeded 10x growth, fueled by agents. Anthropic's efficiency edge (smarter models, leaner staff) couldn't compensate. They now lack servers to serve current models, let alone next-gen, forcing rationing via quotas on paid tokens—tokens users already bought.
Uptime reflects the strain: Claude.ai at 98.8%, API barely 99%, vs. ChatGPT API's 99.9% and Codex's 99.98%. No ".999" precision signals deeper issues.
Opaque Policies Punish Power Users, Sour Trust
To manage overload, Anthropic wields carrot-and-stick tactics targeting the 7% of frontier users (agentic harness runners like OpenClaw, Hermes). Carrots: Spring break 2x usage off-peak (e.g., midnight-2am Pacific, weekends), which worked—shifting async jobs preserved quotas. Sticks followed:
- March 26: Peak-hour (5-11am PT weekdays) session limits tightened; 7% hit caps faster, no specifics on model/token burn rates.
- April 3 (Friday 4pm, Easter weekend): Subscriptions block third-party tools like OpenClaw; switch to API keys, discounted bundles, or extra usage. One-time credit equals plan cost ($20 Pro, $100 Max, $200 Max 20x). > "Our subscriptions weren't built for the usage patterns of these third-party tools."
- Conflicting comms: Public statements clash with employee tweets (e.g., Tariq Shahippar: "Nothing is changing... use Agents SDK for local dev, API for business"). Promises of docs clarity (April 3) unfulfilled by April 22. Is OpenClaw TOS-compliant via Agents SDK? Unclear.
Experimentally, Claude Code vanished from Free/Pro tiers (rolled back), no announcement—just pricing page edits. Logic defies sense: Why block limited-token users from Code, potentially hooking them to upgrade? Policies push manual chat/Code use, not agents.
This opacity erodes trust: Users pay $200/month yet can't fully utilize tokens. Business model assumes underutilization (like car wash subs), but enforcing it via bans feels predatory.
OpenAI Swoops In as White Knight
Anthropic's self-inflicted PR nightmare hands OpenAI free wins. Sam Altman tweets Codex resets for 3M weekly users (promising every million to 10M). Acquired OpenClaw's Peter Steinberger; full sub support, no TOS drama—though Opus edges GPT-4o in agents, reliability trumps.
OpenAI captures excess demand: Anthropic's bans/downtime funnel users. Anthropic's enterprise focus blinded them to consumer/agent boom; now, quota fights alienate devs they need for data flywheel.
Key Takeaways
- Prioritize capex aggression in AI: Demand outpaces conservative forecasts; secure compute now or lose flywheel momentum.
- Communicate policies transparently: Specifics on quota burn, TOS clarity prevent confusion—update docs immediately, don't promise and ghost.
- Embrace agentic usage: Third-party harnesses like OpenClaw drive adoption/data; blocking them kills hooks to paid tiers.
- Monitor uptime religiously: 98.8% vs. 99.9% competitors signals compute gaps—invest to match.
- Shift peak loads: Off-peak incentives work; build async agent support into products.
- For builders: Test OpenAI Codex/OpenClaw as Anthropic hedge; Opus shines but reliability matters.
- Enterprise flywheels need consumer moats: Anthropic's AGI purity ignored agent hype cycle.
"This is the most insane flywheel I've ever seen... probably the most insane flywheel we've ever seen in the history of business." "We're all looking at each other... I still have no idea if what I'm doing with OpenClaw and Claude is actually within their terms of service." "They are doing everything they can to prevent me from using my tokens that I'm paying them for."