The AI compute economy just got its clearest price tag yet: Anthropic is paying SpaceX $1.25B/month — $15B/year — for access to the COLOSSUS II cluster through May 2029, a deal that simultaneously explains Anthropic's path to profitability and cements SpaceX as a major AI infrastructure player overnight. Anthropic is projected to hit ~$500M quarterly profit in Q2 2026, meaning the company is essentially running on razor-thin margins while betting its entire training future on a single non-traditional compute vendor. Meanwhile, Google I/O 2026 dropped 100 announcements including Gemini Omni and Gemini Spark, and OpenAI's general-purpose model autonomously cracked an 80-year-old Erdős conjecture — the kind of week that makes every prior benchmark discussion feel quaint.
SpaceX's S-1 filing reveals Anthropic is paying $1.25B/month for compute access to the COLOSSUS II cluster, locked in through May 2029. The deal emerged from the S-1 disclosure, not an Anthropic announcement, and positions SpaceX as one of the largest compute vendors in AI — ahead of many purpose-built cloud providers.
Why it matters
This is the largest known single compute contract in AI history and signals that frontier training capacity has become so scarce that labs are buying outside hyperscalers entirely. It also puts a hard floor under Anthropic's cost structure just as it approaches profitability.
OpenAI's general-purpose model autonomously solved the planar unit distance problem, an Erdős conjecture open since the 1940s, discovering new geometric constructions that outperform the previously best-known grid-based solutions. OpenAI describes it as "the first time AI has autonomously solved a prominent open problem central to a field of mathematics."
Why it matters
Autonomous mathematical discovery — not assisted proof-checking, but novel construction — is a qualitative capability threshold. This puts concrete distance between frontier models and the "stochastic parrot" framing that still dominates public discourse.
Google announced Gemini Omni (full multimodal flagship), Gemini Spark (a lightweight conversational model competing with OpenAI's o-series), and Gemini 3.5 Flash hitting general availability, across a 100-item announcement slate. Simon Willison notes most launches lacked hands-on availability at announcement time, with many in preview or waitlist status.
Why it matters
The volume signals Google is playing catch-up across every AI product category simultaneously. Gemini 3.5 Flash scoring 76.7% on SimpleBench — 0.2 points behind GPT-5.5 Pro — shows the efficiency tier is genuinely competitive, which matters more for API revenue than flagship benchmarks.
Anthropic is on track to reach profitability in Q2 2026, with approximately $500M in quarterly profit driven by Claude API revenue scaling. This follows a period of aggressive capital raises and comes as the COLOSSUS II compute deal clarifies the cost side of the ledger.
Why it matters
Anthropic becomes the first safety-focused frontier lab to reach profitability, fundamentally changing its negotiating position with investors, regulators, and enterprise customers. It also validates the API-first revenue model over consumer subscriptions.
Railway has repositioned as an agent-native cloud platform — 3M users, 100K weekly signups — with its own data centers and over $200K in monthly spend attributable directly to coding agents. CEO Jake Cooper frames agents as the primary infrastructure consumer, not developers spinning up side projects.
Why it matters
Railway is the clearest early signal that coding agents are generating real, measurable cloud infrastructure revenue at scale. When a mid-sized cloud provider is tracking agent spend as a distinct line item, the agent economy has moved from thesis to billing cycle.
Midjourney publicly stated that its decision to build on TPU infrastructure delayed research progress by approximately one year, and the company regrets not committing exclusively to NVIDIA from the start. The admission came without a detailed technical breakdown but was unambiguous about the cost.
Why it matters
This is a rare public acknowledgment of hardware platform risk from a major AI lab, and it will harden the already-strong case for NVIDIA lock-in across the industry. Any lab currently on a TPU-first or hybrid strategy now has a named cautionary tale to defend against.
Midjourney's TPU regret, Anthropic's COLOSSUS II NVIDIA-based cluster deal, and Railway's agent-native infrastructure buildout all point toward consolidation around NVIDIA hardware as the default training and inference substrate — with alternatives paying a measurable research tax.
Profitability is arriving faster than the market expected
Anthropic approaching $500M quarterly profit in Q2 2026 follows closely on reports of OpenAI and Google Cloud AI divisions turning cash-flow positive, suggesting the frontier lab cost structure is more manageable at scale than the capital-raise pace implied.
Open-weights MoE is becoming the default local inference architecture
Cohere's Command-A-Plus release, Qwen 3.6 35B benchmarks on RTX 5080 at 56 tok/s, and multiple GGUF quantization threads all point to sparse MoE as the practical standard for sub-$4K consumer hardware — with Multi-Token Prediction still failing to deliver its theoretical gains in real deployments.
Signal vs noise
Signal · underreported
The SpaceX S-1 compute disclosure is the most structurally important data point of the week and got almost no mainstream coverage. A non-hyperscaler controlling a $15B/year training contract for a frontier AI lab rewrites assumptions about who controls AI infrastructure. Watch for AWS, Azure, and GCP to respond with aggressive pricing or exclusivity terms.
Noise · overhyped
The Gemini 3.5 Flash vs. GPT-5.5 Pro SimpleBench gap of 0.2 percentage points generated significant engagement but means nothing operationally — SimpleBench scores at this level are within noise margin, and neither model's enterprise use cases hinge on this benchmark.
What to watch
Track whether OpenAI publishes a formal paper on the Erdős unit distance result — peer review and reproducibility will determine whether this is a genuine mathematical milestone or a well-executed search over known construction families.