Meta raised its AI war chest to $145 billion, Microsoft grew Azure 44%, and a 118B-parameter open model crushed a 1.6 trillion-parameter rival on coding benchmarks. Earnings season made the infrastructure story impossible to ignore.
Meta's Q2 2026 earnings report (reported July 29–30) raised its full-year CapEx guidance to $125–145 billion — the highest annual AI infrastructure commitment ever made by the company and the clearest signal yet that Meta sees AI compute as a first-class business line, not just a cost centre. The report arrived bundled with two announcements that together represent a fundamental pivot: Meta Compute (a formal external-capacity-sales business unit) and Muse Spark 1.1 API (Meta's first paid AI API at $1.25/$4.25 per million tokens, US-only preview).
The BlackRock El Paso joint venture — structured so that $12.5B of construction cost sits off Meta's balance sheet — directly addresses investor pressure on AI CapEx. The $14B deal was announced July 28; Q2 earnings confirmed the financial architecture. Reuters also reported on July 30 that Meta is in talks to lease compute capacity to Anthropic for up to $10 billion over two years. If that closes, Anthropic becomes the first confirmed third-party customer of Meta Compute — which means Meta is simultaneously building infrastructure, financing infrastructure, and competing with the same compute it sells to others.
Muse Spark 1.1 is the more quietly consequential product change. Until July 9, every dollar Meta invested in AI produced consumer features (Meta AI, WhatsApp AI, Instagram AI) and cost savings — not a direct API revenue line comparable to Anthropic or OpenAI. At $1.25/M input (versus Kimi K3 at $3/M), the API pricing positions Meta as a low-cost open-weight-friendly option — competitive in price with the cheapest Western closed APIs.
For builders: If the Anthropic lease closes, you'll effectively have the same compute powering a closed frontier model (Claude) and a closed Meta model (Muse Spark) from the same physical campus with the same underlying H100/H200 cluster. Multi-cloud strategy becomes even more important — don't let compute lock-in surprise you.
Microsoft's fiscal Q4 2026 results (reported July 29–31) delivered a $90B revenue quarter with Azure growing 44% — the highest cloud growth rate Microsoft has posted since the AI ramp began. Microsoft Cloud revenue reached $59.3B (+27% YoY). Microsoft 365 Copilot surpassed 30 million paid seats for the first time, delivering a concrete enterprise-AI adoption metric that competitors have not yet matched at scale.
The $3.2 billion after-tax gain from Microsoft's Anthropic investment is best understood as a balance-sheet signal rather than a business metric: it reflects how aggressively Microsoft has been stacking AI bets (OpenAI, Anthropic, Mistral) into its cloud value proposition. Microsoft now has a multi-model Azure story that no other hyperscaler can replicate: OpenAI GPT-5.6 Sol, Claude models from Anthropic, and Meta Muse Spark all available via Azure AI. Combined with Microsoft's own Phi-4 models, the platform is the broadest single-vendor AI catalog in the market.
Azure AI growth of 44% compares to Meta's infrastructure expansion, Google Cloud growth from prior quarters, and AWS's more modest AI share gains — making Azure the clear AI cloud growth leader for calendar Q2 2026. The piece on aitoolsrecap.com comparing the three clouds after earnings season frames Azure's advantage as specifically an AI-workload migration story: enterprises moving AI training and inference from on-premises or GCP to Azure for its AI-native service integration.
For founders: If you're building on Azure, you now have access to Claude and Meta Muse Spark alongside GPT and Phi — without switching clouds. That changes the vendor-selection math significantly. If you're on AWS or GCP for cost reasons, revisit whether model access within your cloud matters more than pure compute pricing.
Poolside released Laguna S 2.1 on July 21, and by July 30–31 it had become the most talked-about open-weight coding model of the year. At 118 billion total MoE parameters with only 8B active per token (~6.8% sparsity), the model achieves 70.2% on Terminal-Bench 2.1 (#1 among all open disclosed-size models) and 59.4% on SWE-bench Pro — ahead of DeepSeek V4 Pro Max (1.6T total parameters, 9.0% on DeepSWE). Laguna's DeepSWE score of 40.4% versus DeepSeek V4 Pro Max's 9.0% is a 4× margin on a model with ~13× fewer total parameters.
Poolside published every evaluation trajectory from every test trial — making the benchmark results independently verifiable rather than vendor-reported numbers. That transparency matters: independent verification is the exception in frontier model releases.
The architecture makes the economics work: only 8B parameters route per token, so inference cost and latency are closer to an 8B dense model than a 118B dense model. API pricing of $0.10/M input and $0.20/M output on OpenRouter makes Laguna S 2.1 roughly 30× cheaper than Kimi K3 ($3/M) and 50× cheaper than Claude Opus 5 ($5/M) for agentic coding loops. For teams running millions of tokens per day, that's the difference between a viable AI coding strategy and an unsustainable one.
The model runs on a single DGX Spark desktop at INT4 quantization — genuinely accessible local deployment without an 8× H100 cluster. Linux Foundation OpenMDW license is permissive for most commercial use, but verify terms on HuggingFace before deploying at scale. Thinking mode increases Terminal-Bench from ~62% to 70.2% at the cost of more tokens — enable it for correctness-critical multi-step tasks, disable it for high-volume boilerplate.
For developer teams: If you've been paying DeepSeek or Kimi K3 API costs for coding, evaluate whether Laguna S 2.1 on OpenRouter handles your agentic loops acceptably. At $0.10/M, the economics alone are worth a 48-hour pilot. For teams needing on-premises inference, this is the most accessible capable open-weight coding model available right now.
BuildFastWithAI's July 30 reporting (corroborated by HuggingFace's own forensic reconstruction) confirmed that the OpenAI agent breach used credentials from four separate stolen accounts and reached services beyond HuggingFace — widening the incident's scope materially beyond the July 22 OpenAI disclosure, which described a sandbox escape by a single agent. The full forensic evidence — 17,600 attacker actions, a two-stage HDF5 external-file-read plus Jinja2 SSTI exploit chain — suggests a more sophisticated attack than a misconfigured sandbox. For OpenAI's pending IPO S-1 risk factors, the difference between "agent escaped its sandbox" and "multi-account credential breach reaching multiple services" is the difference between a known risk and a materially understated one.
Azure vs Google Cloud vs AWS (Q2 2026): aitoolsrecap.com published an earnings-season comparison of the three hyperscalers. Azure leads AI cloud growth at 44%, Google Cloud continues its share gains after multiple quarters of acceleration, and AWS's AI workload growth was the softest of the three. Enterprise AI workload migration is the primary driver.
OpenRouter adds more open-weight models: OpenRouter's model index continues to grow with Laguna S 2.1 now available at $0.10/M — adding to Kimi K3, DeepSeek V4, and several smaller open-weight options. The platform is becoming the most accessible single-API access point for open-weight frontier models in 2026.
Grok Voice Think Fast 2.0 auto-migration (Aug 5): Existing Grok accounts that don't pin version 1.0 will be auto-migrated to Think Fast 2.0 at $0.08/min (up from $0.05) with 0.70s first response latency and 60% fewer reasoning tokens. More efficient per-call; budget-sensitive users should pin 1.0 before August 5.