Power constraints now the critical bottleneck for data center expansion through 2030

Enterprise token budgeting is now standard, but lacks consensus on spending caps—and tokenmaxxing headlines were overblown

SemiAnalysis surveyed 50+ enterprises and found widespread adoption of token limits ranging from $250–$2,000+ monthly per employee, with no industry convergence on a specific number. The widely reported token consumption blowouts at Meta (60T tokens/month) and Uber stemmed from misaligned internal incentives—Meta employees competed for "Token Legend" rankings by running agents for hours purely to burn tokens—rather than high-ROI use cases. In contrast, companies with mature AI adoption (Amazon, Stripe, Workday) justify higher spend through measurable productivity gains: Amazon cut its recruiting cycle from 6–9 months to 3–4 months using AI tools for interview notes and reporting. Most non-tech-forward Fortune 500s still spend under $100 per employee annually on AI, suggesting the enterprise adoption S-curve has significant runway ahead.

Coding use cases drive 70%+ of Anthropic and OpenAI ARR; TaaS providers projected at $4B+ ARR

SemiAnalysis's Tokenomics Model reveals extreme concentration in API revenue: 99th percentile customers spend ~$90K/year per employee versus a median of $136, with the 90th+ percentile accounting for most revenue growth. The model estimates that over 70% of combined Anthropic and OpenAI ARR today comes from coding use cases (higher for Anthropic given its 90%+ B2B mix versus OpenAI's 60%), with application-layer companies like Cursor, Loveable, and GitHub Copilot capturing significant value. Token-as-a-Service providers (Together, Fireworks, Baseten) are projected at $4B+ ARR, indicating massive demand for both frontier and open-source model endpoints as enterprises move beyond initial adoption into white-collar knowledge work verticals.

Cloud agents emerging as standard infrastructure across OpenAI, Anthropic, and Cursor

OpenAI acquired Ona (formerly Gitpod) to enable long-running agents in sandboxed cloud environments, while Anthropic built Claude Managed Agents as a six-month infrastructure project and Cursor launched cloud agents with iOS PWA support. Cloud agents address practical limitations of local execution—CPU heat, setup complexity, inability to run long tasks—and are enabled by improved coding models, larger context windows, and abundant GPU capacity. Cursor's implementation allows agents to clone repos, run tests, interact with databases and APIs via MCP servers, and produce merge-ready PRs with artifacts; agents can be triggered from Cursor Web, Desktop, Slack, GitHub, Bitbucket, Linear, or API. This shift reflects a maturation phase where infrastructure and economics, not just capability, are driving product strategy.

US model export restrictions create uncertainty for frontier labs, incentivizing open-source and non-US alternatives

The US government relaxed restrictions on Anthropic's latest models to allow access for ~100 US companies and agencies, while OpenAI faced similar constraints. This arbitrary policy creates risk for companies building around models that could be shut down, incentivizing more open-source development, edge deployment, and potentially frontier labs outside the US or China—though DeepSeek's $7.5B funding round shows the cost of staying competitive is prohibitive.

Anthropic lifts export controls on Fable 5 and Mythos 5 after Commerce Department approval

The Department of Commerce cleared Anthropic's models for unrestricted access after a two-week review, with the company restoring availability starting July 2. This signals government alignment on frontier AI development and removes a significant distribution constraint for Anthropic's product lineup.

Etched AI exits stealth with $800M raised and $1B+ customer contracts; OpenAI claims halved inference costs

Etched AI raised $800M from investors including Jane Street and a TSMC-linked VC, with $1B+ in signed customer contracts and first racks shipping summer 2026. The company claims successful A0 tapeout and SOTA throughput, latency, and power efficiency on inference workloads, with a technical approach centered on low-voltage inference and cluster-scale memory for long-context workloads. Simultaneously, OpenAI engineers claim to have halved inference costs, and Omen AI raised $31M for data center optimization. These signals point to intensifying competition on inference economics as the bottleneck shifts from model capability to deployment efficiency.

Anthropic's Sonnet 5 costs more per task than Opus 4.8 despite lower list pricing due to token verbosity

Sonnet 5 uses ~40% more output tokens and 3–6x more agentic turns than Sonnet 4.6, resulting in $2.29 per Intelligence Index task—roughly 15% more expensive than Opus 4.8 at standard pricing. At 69k output tokens per task on average, Sonnet 5 is one of the highest token users, behind only GPT-5.4 mini and nano. This tokenizer/verbosity issue undermines the headline cost-performance narrative and explains why power users and evaluators found the launch underwhelming despite meaningful capability gains over Sonnet 4.6.

Engineering work increasingly focused on optimizing agent efficiency and cost-per-token economics

Platform teams at Anthropic and Cursor are shifting focus toward building environments where agents execute more efficiently, while companies like Coinbase are aggressively optimizing spend-per-token as AI usage costs become material. This reflects a maturation phase where infrastructure and economics, not just capability, are driving product strategy.

Global Data Center Trends 2026

CBRE published a comprehensive report on global data center market dynamics through Q1 2026. Main takeaways:

• North American inventory grew 33% YoY despite record-low vacancy (Northern Virginia 0.3%, Atlanta 1%), indicating demand outpacing supply across all major hubs
• Power availability and grid constraints are now the binding constraint on development timelines in established markets (Northern Virginia, Chicago, Frankfurt, London), with emerging regions (Querétaro +450% inventory, Johor, West Texas) gaining traction
• Rental rates rising to unprecedented levels: Chicago $200-$230/kW/month, Frankfurt $235-$265/kW/month, Singapore $330-$475/kW/month; Frankfurt saw highest European rates due to limited inventory
• Net absorption surged 34% YoY in North America to 2,236.2 MW in Q1, with Northern Virginia setting all-time high of 1,148.3 MW driven by hyperscalers and AI startups seeking continuous power
• Available capacity at all-time lows despite massive inventory additions (Dallas added 379.9 MW inventory but only 1.4 MW available capacity), signaling power procurement constraints are the critical bottleneck through 2030

SemiAnalysis AI Tokenomics Model

SemiAnalysis published a detailed demand-side AI economics model connecting hardware inputs to software outputs across the AI stack. Main takeaways:

• Bottoms-up token throughput forecasts by hardware system (GB200 NVL72, VR144, TPU v7, Trainium 3), model architecture (GPT 5, Sonnet 4, DeepSeek V3), and workload type (coding, chat, agentic)
• Tracks AI compute supply/demand across hyperscalers, foundation labs (OpenAI, Anthropic, DeepSeek), and neoclouds (Coreweave, Nebius, Crusoe) with ClusterMAX quality ratings
• Models disruption of traditional SaaS by token-consumption companies, affecting Salesforce, Workday, Adobe, ServiceNow, Microsoft and others
• Calculates ROIC of AI deployments and derives aggregate hardware demand forecasts from usage/revenue growth across inference and training
• Provides addressable market analysis of token economy including Google AI Overviews, ChatGPT APIs, and token-consuming software (Cursor, Harvey, Perplexity)

Forward Deployed Engineers and the future of software engineering

Latent Space published an interview with Natalie Meurer, Head of Agent Engineering at Sierra, on the evolving definition and role of forward deployed engineers in AI. Main takeaways:

• Sierra intentionally uses "agent engineer" instead of "forward deployed engineer" to describe a specific technical role combining systems integration, agent development, and customer operations understanding, rather than a vague accountability-based title.
• Most customer-specific work at Sierra happens at the orchestration layer above models, with enterprises reaching production in 40-60 days despite requiring custom integration across APIs, systems, SOPs, and brand tone.
• Product engineering and forward deployed engineering are converging as code becomes cheaper to author; the best practitioners in each role increasingly need customer-facing and product-building skills respectively.
• Agent development for enterprise customer service differs fundamentally from other agentic use cases: solving the same problem 100,000 times weekly with consistent processes, rather than novel problem-solving, creating persistent demand for customer-facing engineers who translate business rules into systems.