Anthropic just hit $30 billion in annual recurring revenue, surpassing OpenAI to become the highest-grossing AI company in the world. The company grew 30x in 15 months, doubled its $1M+ enterprise customers to over 1,000 in just two months, and positioned itself as the enterprise AI leader. Then, in late March 2026, developers started noticing something wrong with Claude Code, Anthropic's flagship AI coding tool. The model was making more mistakes, skipping critical steps, and becoming what one AMD senior director called "unusable for complex engineering tasks."
What happened? According to widespread user reports and an internal memo leaked from OpenAI, Anthropic appears to be facing acute compute capacity constraints—a victim of its own explosive growth. The company quietly reduced Claude's default "effort level" to conserve tokens and computing resources, triggering a backlash from enterprise users who rely on the tool for mission-critical work.
For CIOs, CTOs, and VPs of Engineering evaluating AI vendors, this situation is a case study in how infrastructure capacity can become a hidden risk factor—especially when you're spending $1M+ per year on AI tools. Here's what enterprise leaders need to understand about compute constraints, token economics, vendor transparency, and how to evaluate AI infrastructure resilience when selecting enterprise AI partners.
The Performance Collapse: What Users Are Seeing
Claude Code, Anthropic's autonomous coding agent, became one of the company's fastest-growing products after launching in early 2025. Enterprise engineering teams adopted it for complex, multistep coding tasks—code reviews, refactoring, debugging, and feature development. Unlike ChatGPT or GitHub Copilot, Claude Code operates as a command-line agent that reads, writes, and executes code autonomously within a developer's environment.
Then, starting in late February 2026, users noticed a dramatic shift in Claude's behavior. A GitHub analysis posted by Stella Laurenzo, a senior director of AI at AMD, documented the regression in detail. The analysis showed that Claude moved from a "research-first" approach—reading multiple files and gathering context before making changes—to a more direct "edit-first" style. The model read less context before acting, made more mistakes, and required significantly more user intervention.
Laurenzo's findings included:
- Reduced context gathering: Claude stopped reading relevant files before making edits
- Premature stopping: The model quit tasks early, avoiding responsibility
- Increased mistakes: More errors requiring manual correction
- Over-asking permission: Unnecessary user intervention requests
"Claude has regressed to the point [that] it cannot be trusted to perform complex engineering," Laurenzo wrote in the widely shared analysis.
Dimitris Papailiopoulos, a principal research manager at Microsoft, echoed the frustration on X: "I've had incredibly frustrating sessions with Claude Code the past two weeks. I set effort to max, yet it's extremely sloppy, ignores instructions, and repeats mistakes."
This wasn't just hobbyist complaints—these were senior engineering leaders at AMD and Microsoft reporting production-level tool failures. For enterprise teams relying on Claude Code for daily workflows, the degradation had immediate productivity implications.
The Token Economics Problem: Why Anthropic Made the Change
Boris Cherny, the Anthropic executive who leads Claude Code, responded to user complaints by explaining the company had reduced Claude's default "effort level" to "medium" in early March 2026. The change was made, he said, in response to user feedback that Claude was consuming too many tokens per task.
Here's what that means in practice:
Tokens are the units of data an AI model processes—more tokens = more compute = higher costs. When Claude uses a higher "effort level," it spends more time reasoning through a problem, gathering context, and validating its approach before acting. This produces better results but consumes more computing power.
Anthropic's solution was to introduce "adaptive thinking" in February 2026—allowing the model to decide how much reasoning to apply to each task rather than using a fixed budget. Then, in March, the company shifted the default setting down to "medium effort."
The problem? Users weren't clearly notified of the change, and the impact was dramatic. Claude Code users can manually adjust effort levels, but users paying for Pro versions of Claude Cowork or the desktop version of Claude cannot currently change the default. Cherny said the company would test "defaulting Teams and Enterprise users to high effort" going forward, acknowledging the issue affected production workflows.
For enterprise buyers, the token economics question is critical: How much are you willing to pay for higher-quality outputs, and does your vendor's pricing model align with your actual needs? If Anthropic is optimizing for token efficiency at the expense of output quality, what does that mean for teams relying on Claude for complex, high-stakes work?
The Compute Capacity Question: Is Anthropic Running Out of GPUs?
OpenAI's Chief Revenue Officer didn't mince words in an internal memo reported by CNBC: Anthropic made a "strategic misstep" by not securing enough compute capacity and is "operating on a meaningfully smaller curve" than competitors.
The evidence supporting this claim is circumstantial but compelling:
1. Usage limits during peak hours: Anthropic introduced stricter usage caps in March 2026, drawing user complaints. When Anthropic capped usage, OpenAI immediately announced it would double its limits—a competitive jab highlighting the contrast.
2. Recent outages: As Claude adoption surged, the platform suffered multiple outages, suggesting infrastructure strain.
3. New model announcement without release: Anthropic announced Mythos, a more capable (and more expensive to run) model, but stressed it's not releasing it to the general public yet due to "security concerns." Some industry observers questioned whether Anthropic lacks sufficient compute to support a broad Mythos rollout.
4. Compute capacity deals: While Anthropic has secured significant infrastructure agreements—1 gigawatt of Google TPU capacity announced in October 2025, with multiple gigawatts coming online starting in 2027—the timing suggests current capacity may be strained.
Compare that to OpenAI's recent deal with AWS: 2 gigawatts of Trainium computing capacity, a massive commitment that gives OpenAI immediate access to next-generation infrastructure.
For enterprise leaders, the question isn't just "Is Anthropic running out of compute?" It's "What happens to my workflows when my AI vendor hits capacity constraints?" Reduced performance, usage limits, and outages aren't just technical inconveniences—they're productivity blockers for teams spending $1M+ per year on AI tools.
The Transparency Problem: Why Disclosure Matters
Anthropic has built its brand on being more transparent and user-aligned than competitors. The company publicly fought the U.S. Department of Defense over ethical use restrictions, insisting on contractual language prohibiting use in lethal autonomous weapons or mass surveillance. That principled stance earned significant goodwill and drove a wave of new users switching from OpenAI's ChatGPT.
But the Claude Code performance degradation exposed a transparency gap. Cherny said the effort-level change was flagged via a pop-up within the Claude Code interface, but many users said they didn't see it or didn't understand the implications. The backlash suggests the disclosure was insufficient for the scale of impact on production workflows.
For enterprise buyers, vendor transparency is a risk management tool. When you're integrating AI into critical business processes, you need to know:
- When performance changes are coming
- Why they're being made
- How they'll affect your workflows
- Whether you can opt out or configure differently
Without clear communication, enterprises are left guessing whether performance issues are bugs, capacity problems, or intentional optimizations. That uncertainty undermines trust and makes long-term planning difficult.
What This Means for Enterprise AI Vendor Selection
Anthropic's compute crunch offers five key lessons for CIOs, CTOs, and VPs evaluating AI vendors:
1. Compute Capacity Is a Vendor Selection Criterion
Don't just evaluate model performance—evaluate infrastructure commitments. Ask vendors:
- What computing capacity do you have under contract?
- When does new capacity come online?
- What's your plan for scaling infrastructure to match demand?
- What happens to performance or availability when you hit capacity constraints?
Vendors with multi-gigawatt commitments (like OpenAI's AWS deal) have more headroom for growth. Vendors relying on future capacity (like Anthropic's 2027 Google TPU expansion) may face near-term constraints.
2. Token Pricing Models Matter
Understand how your vendor charges for compute and whether pricing incentives align with your quality expectations. If you're paying per token, does higher quality cost more? Can you lock in a fixed effort level? Are there hidden costs when models consume more compute than expected?
Anthropic's token optimization strategy makes financial sense for the company but created friction for users expecting consistent performance. Enterprise buyers should negotiate SLAs that guarantee minimum performance levels regardless of vendor optimization efforts.
3. Multi-Cloud Distribution Is a Strategic Advantage
Anthropic's multi-cloud strategy (AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry) gives it more infrastructure flexibility than OpenAI (Azure-only) or Google Gemini (Google Cloud-only). If one cloud provider hits capacity constraints, Anthropic can shift workloads.
For enterprise buyers, vendors with multi-cloud distribution offer:
- Redundancy: Less risk of single-cloud outages
- Pricing leverage: Ability to negotiate across providers
- Flexibility: Match workloads to optimal infrastructure
4. Transparency and Change Management Are Non-Negotiable
Require vendors to disclose performance changes, capacity constraints, and optimization strategies that could affect production workflows. Build this into contracts:
- Advance notice for material performance changes
- Changelog visibility for model updates
- Capacity constraint disclosures
- Escalation paths for critical issues
5. Growth Can Be a Double-Edged Sword
Anthropic's 30x revenue growth in 15 months is impressive—but it also created infrastructure strain. Rapid adoption can outpace capacity planning, leaving enterprise customers dealing with performance degradation or usage limits.
When evaluating fast-growing AI vendors, ask:
- How are you managing infrastructure scaling to match demand?
- What's your track record on uptime and availability during growth phases?
- Do you have buffer capacity for unexpected demand spikes?
The Bigger Picture: AI Infrastructure Is the New Enterprise Risk
Anthropic is projected to spend $30 billion per year on AI training by 2030—a quarter of what OpenAI expects to spend ($125 billion), yet Anthropic is generating more revenue. That efficiency advantage is real, but it doesn't eliminate capacity constraints. It just means Anthropic is doing more with less.
For enterprise AI buyers, the lesson is clear: infrastructure capacity is no longer just an IT concern—it's a strategic risk factor. When your sales team, engineering team, customer support, legal department, or finance operations depend on AI tools, vendor compute constraints directly affect your business operations.
The questions to ask aren't just about model capabilities—they're about infrastructure resilience:
- Can your vendor scale infrastructure as fast as you scale usage?
- What happens to performance when capacity is strained?
- How transparent is your vendor about capacity planning?
- Do you have contractual protections for performance degradation?
What Anthropic Needs to Do Next
Anthropic has a revenue and efficiency advantage over OpenAI, but user trust is fragile. To maintain its enterprise leadership position, the company needs to:
1. Restore default performance for enterprise customers: Defaulting Teams and Enterprise users to "high effort" (as Cherny proposed) is a good start, but users need control over configuration.
2. Improve transparency on capacity and performance: Provide clear disclosures when infrastructure constraints affect performance, with advance notice and mitigation options.
3. Accelerate infrastructure expansion: The Google TPU capacity coming online in 2027 is important, but Anthropic needs near-term solutions to handle current demand.
4. Establish clear SLAs for enterprise customers: Performance guarantees, uptime commitments, and escalation paths for critical issues.
5. Rebuild trust through consistent communication: The user backlash wasn't just about performance—it was about feeling blindsided by undisclosed changes.
The Takeaway: Infrastructure Is the Hidden Risk in Enterprise AI Contracts
Anthropic's compute crunch is a reminder that AI infrastructure is as critical as model performance when evaluating enterprise AI vendors. Revenue growth, model capabilities, and brand reputation matter—but if your vendor can't scale infrastructure to match demand, your workflows suffer.
For CIOs and CTOs spending $1M+ per year on AI tools, the due diligence checklist now includes:
- ✅ Compute capacity commitments
- ✅ Multi-cloud distribution strategy
- ✅ Token pricing transparency
- ✅ Performance SLAs and guarantees
- ✅ Change management and disclosure practices
- ✅ Infrastructure scaling roadmap
Anthropic's explosive growth is impressive—but growth without infrastructure resilience is a risk factor. Enterprise buyers need to evaluate vendors not just on what they can do today, but on whether their infrastructure can scale to meet tomorrow's demand.
When you're betting your engineering productivity, customer operations, or business intelligence on AI tools, vendor compute capacity isn't a technical detail—it's a business continuity question.
Continue Reading:
