AI Weekly Brief
Anthropic scales compute and services
The $1.5B Anthropic/Blackstone/Goldman services firm turns forward-deployed AI engineering into an equity-backed business model.
TL;DR
- Anthropic secured all 300 MW and 220,000+ NVIDIA GPUs of SpaceX's Colossus 1 data center and launched a $1.5B enterprise AI services firm with Blackstone, Goldman Sachs, and Hellman & Friedman, two structural moves that reshape how frontier models reach enterprise customers.
- OpenAI shipped three voice and audio API models led by GPT-Realtime-2, the first voice model built on GPT-5-class reasoning, and made GPT-5.5 Instant the ChatGPT default with 52.5% fewer hallucinated claims than its predecessor.
- Google DeepMind's AI Co-Mathematician, a stateful multi-agent research workspace, set a new state of the art of 48% on FrontierMath Tier 4, as AlphaEvolve marked a year of production impact across Google infrastructure, DNA sequencing, and drug discovery.
- ScaleLogic formalized RL training compute as a power law of reasoning depth, giving post-training its first principled scaling framework analogous to pre-training scaling laws.
- OpenAI expanded its ChatGPT ads pilot to the UK, Mexico, Brazil, Japan, and South Korea, turning a US-only February test into a global monetization track.
Editor’s Pick
Traditional technology consulting runs on labor arbitrage and methodology frameworks that do not transfer cleanly to AI-era transformation, where the bottleneck is the engineering capacity to adapt frontier models to proprietary workflows at production fidelity. Blackstone's own portfolio companies grew LLM spending 15x in a single year, constrained not by model capability but by implementation capacity, which is precisely the bottleneck this firm is designed to resolve.
The structure is notable: the PE firms contribute both capital and a captive customer base (Blackstone alone holds 275 portfolio companies), giving the firm's forward-deployed engineers a ready-made pipeline. Anthropic gains recurring revenue, deep operational knowledge of portfolio workflows, and preferential mid-market distribution, while the equity investors gain returns correlated with AI adoption across their portfolios and a first-mover position in a new professional-services segment.
The competitive implication is a divergence from OpenAI's GSI channel model (Accenture, Capgemini, Cognizant, and others with Codex), which routes implementation through consulting firms that keep the client relationship. Anthropic's model retains both the relationship and the workflow knowledge, generating equity value and recurring revenue per operating company rather than per API call. That divergence will likely show up in enterprise ARR before it shows up in benchmark comparisons, making it one of the most durable competitive asymmetries introduced this week.
Model Releases
OpenAI's GPT-5.5 Instant. The new ChatGPT default for all users makes 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes medical, legal, and financial prompts, and adds personalization from past chats, files, and Gmail with new memory sources showing which context shaped each answer.
OpenAI's GPT-5.5-Cyber. OpenAI scaled its Trusted Access for Cyber program to GPT-5.5 with a cybersecurity-specialized variant gated by identity verification, mirroring Anthropic's Project Glasswing model, and added a $10M Cybersecurity Grant Program in API credits for defensive research teams.
ByteDance's Doubao-Seed-2.0-lite. The first omni-modal model in ByteDance's Doubao family natively understands audio, video, images, and text, takes up to 25 hours of audio input, and claims 19-language translation ahead of Gemini, with a lighter Doubao-Seed-2.0-mini for lower latency.
AI Tools & Products
OpenAI's GPT-Realtime-2 and two audio models. GPT-Realtime-2 is the first voice model with GPT-5-class reasoning and multi-step tool use inside live calls ($32/$64 per 1M audio input/output tokens), joined by GPT-Realtime-Translate (70+ input to 13 output languages, $0.034/min) and streaming GPT-Realtime-Whisper ($0.017/min).
Anthropic launches Claude Managed Agents. Launched at Code with Claude, the hosted service runs sandboxed autonomous agent sessions on Anthropic infrastructure, with Multiagent Orchestration of up to 25 parallel threads (public beta) and long-horizon Dreaming (research preview), as the Claude Code SDK became the Claude Agent SDK.
Anthropic's Agents for Financial Services. Ten new Cowork and Claude Code plugins for financial services and insurance, plus Microsoft 365 integrations, enterprise data connectors, and an MCP app built for regulated, data-residency-sensitive deployments, extending Anthropic's sector-vertical push.
ChatGPT ads expand to five markets. After early US, Canada, and ANZ tests showed no degradation in trust metrics, OpenAI extended its sponsored-labeled ads pilot to the UK, Mexico, Brazil, Japan, and South Korea, with advertisers receiving only aggregate performance data.
xAI adds Grok Connectors and custom voices. xAI added deep app integrations to Grok Web, a higher-realism Grok Imagine Quality Mode API, and Custom Voices with a Voice Library for cloning brand voices from short recordings.
Research Breakthroughs
Google DeepMind's AI Co-Mathematician. A stateful multi-agent workbench on Gemini 3.1 Pro runs parallel workstreams for ideation, literature search, computation, and proving, scoring a state-of-the-art 48% on FrontierMath Tier 4 (vs. 19% for the base model) and solving three problems no AI had solved.
AlphaEvolve's first year in production. Google DeepMind's evolutionary coding agent cut Gemini training time 1.5% and simplified accelerator circuits and data center scheduling inside Google, and was applied externally to DNA sequencing error correction, power grid simulation, molecular simulation, and drug discovery.
ScaleLogic finds a power law for RL compute. RL training compute scales as a power of reasoning depth (R² > 0.99), with the exponent rising from 1.04 to 2.60 as logical expressiveness grows; easy-to-hard curricula improve efficiency and lift downstream math and reasoning benchmarks by up to 10.66 points.
Apple's TIDE embedding memory. A plug-in that injects context-free token vectors at every transformer layer through a depth-conditioned router, targeting rare-token undertraining and contextual collapse in small models without modifying the backbone.
Fudan's Balanced Aggregation fixes GRPO bias. OpenMOSS identifies sign-length coupling and length downweighting as twin aggregation biases in GRPO and fixes both with a parameter-free drop-in, improving stability and scores for Qwen2.5-Math-7B and Qwen3-1.7B across six reasoning and coding benchmarks.
Industry & Ecosystem
Anthropic takes all of SpaceX's Colossus 1. Anthropic gained the full 300+ MW, 220,000+ NVIDIA GPU capacity of SpaceX's Memphis data center, doubling Claude Code rate limits and raising Opus API limits, and both sides signaled interest in multiple gigawatts of orbital compute.
Anthropic, Blackstone, and Goldman form $1.5B services firm. The AI-native services firm, funded by Blackstone, Hellman & Friedman, and Anthropic (~$300M each) and Goldman Sachs (~$150M), embeds Anthropic engineers with forward-deployed teams at mid-market companies; Blackstone portfolio LLM spend grew 15x in a year.
Trends & Predictions
Frontier labs secure compute beyond the traditional hyperscalers
Compute and custom siliconStrengthenedContinues fromApr 25
Anthropic took all of SpaceX's Colossus 1, 300+ MW and 220,000+ NVIDIA GPUs outside its hyperscaler deals, doubling Claude Code rate limits, and both sides floated multiple gigawatts of orbital compute.
- Prediction (Nov 2026): At least two other frontier labs announce compute partnerships with non-traditional providers such as defense contractors, energy companies, or satellite operators as hyperscaler capacity approaches strategic saturation for the top labs.
- Resolves: Yes if by November 8, 2026 at least two frontier labs other than Anthropic (such as OpenAI, Google DeepMind, xAI, Meta, or Mistral) announce compute capacity agreements with a provider that is not a hyperscaler or GPU cloud, such as an energy company, defense contractor, or satellite operator, per the lab's or partner's press release.
Anthropic, Blackstone, and Hellman & Friedman (about $300M each) and Goldman Sachs (about $150M) formed a $1.5B services firm that embeds Anthropic engineers at mid-market companies, betting on demand like the 15x one-year growth in Blackstone portfolio LLM spend.
- Prediction (Nov 2026): OpenAI, Cohere, and at least one more frontier lab announce formally capitalized enterprise services vehicles or SI channel arrangements with equity-like alignment.
- Resolves: Yes if by November 8, 2026 OpenAI, Cohere, and at least one other frontier lab besides Anthropic each announce a separately capitalized enterprise services vehicle or a systems-integrator partnership involving an equity stake, per their press releases.
Based on Anthropic, Blackstone, and Goldman form $1.5B services firm
Google DeepMind's AI Co-Mathematician scored 48% on FrontierMath Tier 4 versus 19% for its base model and solved three problems no AI had solved, while AlphaEvolve's first production year cut Gemini training time 1.5% and reached drug discovery and DNA sequencing.
- Prediction (Nov 2026): A peer-reviewed mathematical or formal-methods result with AI as a primary algorithmic contributor is accepted at a top venue such as STOC, FOCS, or Annals of Mathematics, setting a precedent for AI attribution.
- Resolves: Yes if by November 8, 2026 STOC, FOCS, or the Annals of Mathematics accepts a paper that credits an AI system as a primary contributor to its main result, as shown in the paper or the venue's accepted-papers list.
Based on Google DeepMind's AI Co-MathematicianAlphaEvolve's first year in production
Voice AI splits into commodity APIs and reasoning-grade agents
Voice and multimodal interfacesStrengthenedRevivesApr 19
OpenAI priced GPT-Realtime-2, the first voice model with GPT-5-class reasoning and live tool use, at $32/$64 per 1M audio tokens, while its streaming Whisper at $0.017/min and xAI's Custom Voices filled the commodity tier.
- Prediction (Sep 2026): Voice API pricing splits into commodity transcription and TTS below $0.02/min and reasoning-grade voice above $0.30/min for multi-step tool use in live calls, with the premium sustained by agentic capability rather than audio quality.
- Resolves: Yes if on September 30, 2026 the pricing pages of at least two providers (such as OpenAI, Google, xAI, or ElevenLabs) list a reasoning-grade real-time voice model costing more than $0.30 per minute alongside transcription or TTS priced below $0.02 per minute.
Based on OpenAI's GPT-Realtime-2 and two audio modelsxAI adds Grok Connectors and custom voices
Agent post-training becomes predictable engineering
Agent training and memoryStrengthenedRevivesApr 19
ScaleLogic found RL training compute scales as a power of reasoning depth (R² > 0.99), with easy-to-hard curricula adding up to 10.66 benchmark points, and Fudan's Balanced Aggregation fixed two GRPO aggregation biases with a parameter-free drop-in.
- Prediction (Nov 2026): RL compute scaling laws appear as standard methodology sections in frontier-lab technical reports, replacing empirically tuned, undisclosed RL training schedules.
- Resolves: Yes if by November 8, 2026 technical reports or system cards from at least two frontier labs (such as OpenAI, Google DeepMind, Anthropic, Meta, DeepSeek, or Alibaba Qwen) include a fitted scaling law relating RL post-training compute to model performance.
Based on ScaleLogic finds a power law for RL computeFudan's Balanced Aggregation fixes GRPO bias