AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #22

Compute and Capability

Wednesday, May 6, 2026 · 3 Developments · 2 Perspectives · 2 Case Studies · 1 Resource

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

Two threads converged this morning. Anthropic spent today's Code with Claude developer conference in San Francisco resolving the constraint that has shadowed every Claude rollout since January — capacity — while OpenAI used the same 24 hours to push back on the default-model and compute fronts. The frameworks for thinking through what any of this means for a growth-stage business also got noticeably sharper.

The advisor takeaway: capacity has stopped being the binding constraint, monetization frameworks are catching up, and the case for treating AI strategy as architecture rather than feature work is now harder to dismiss.

Developments

Technology

May 6: Anthropic signs SpaceX for full Colossus 1 capacity, doubles Claude Code limits, ships Cowork plug-ins

Anthropic opened today's Code with Claude developer conference in San Francisco by announcing it has signed a deal to use the entire compute capacity of SpaceX's Colossus 1 data center in Memphis — more than 300 megawatts and over 220,000 Nvidia GPUs coming online within the month. The Pro and Max usage caps that have frustrated Claude subscribers for months were lifted at peak hours, and Claude Code's five-hour rate limits doubled effective today across Pro, Max, Team, and Enterprise plans. The deal sits alongside earlier Amazon, Google-Broadcom, and Microsoft-Nvidia capacity commitments and includes stated interest in eventual orbital compute. The same announcement bundle includes ten new Cowork and Claude Code plug-ins, Microsoft 365 integrations, and an MCP app for financial services and insurance.

So what

Capacity stops being the binding constraint on Anthropic deployment, which directly affects every growth company evaluating Claude in production over the next two quarters. The implication for procurement is straightforward: vendors should be pressed on whether contracted throughput now rises in line with these capacity additions, and engineering leaders should retest workloads that were rate-limited in March and April. The strategic implication is more interesting: Anthropic just absorbed xAI's Memphis investment as raw compute supply, which says something about how the model layer is consolidating capacity wherever it sits. Worth raising with infrastructure and AI leads this week.

Hype check

The 220,000 GPU figure and "within the month" timing come from Anthropic's own announcement; reporters have not independently audited migration timelines, and historical large-cluster transitions have run weeks behind announced schedules.

Confidence: Established fact — primary announcements from both Anthropic and Musk plus same-day reporting from CNBC, Reuters, Bloomberg, and Inc.

Go deeper Anthropic announcement · CNBC · The Neuron

Product & Technology

May 5: Anthropic ships ten ready-to-run finance agent templates with Microsoft 365 integration

At an event in New York yesterday, Anthropic released ten production-grade Claude agent templates targeting the workflows that consume junior banker and analyst hours: pitch builder, meeting preparer, earnings reviewer, model builder, market researcher, valuation reviewer, GL reconciler, month-end closer, statement auditor, and KYC screener. Each template ships either as a Cowork or Claude Code plug-in, or as a cookbook for Claude Managed Agents with audit logs and credential vaults. Claude now operates inside Excel, PowerPoint, and Word via Microsoft 365 add-ins, with Outlook arriving soon, and connectors expanded to include Dun & Bradstreet, Verisk, IBISWorld, Guidepoint, Third Bridge, SS&C IntraLinks, Fiscal AI, and Financial Modeling Prep. Moody's separately launched an MCP app covering credit ratings on more than 600 million entities. Claude Opus 4.7 leads Vals AI's Finance Agent benchmark at 64.37%.

So what

For growth-PE portfolios, the pitchbook-to-month-end-close stack just collapsed from a multi-week build into a same-week deployment. Operating partners and CFOs at portfolio companies should pull the templates from the Anthropic financial services marketplace this week and run a sandbox comparison against the labor those workflows currently consume. The longer-term question is which incumbent vendors lose pricing power: FactSet, S&P Capital IQ, and Pitchbook are now reachable through governed Claude connectors, which changes the renegotiation context. The 64.37% benchmark also implies meaningful human-in-the-loop review remains non-negotiable; treat these as productivity multipliers, not replacements.

Hype check

The 64.37% benchmark score is a leading number relative to peers, but a 35% failure rate is meaningful in regulated workflows where audit and accuracy requirements are tight; Anthropic's own framing emphasizes human review for that reason.

Confidence: Established fact — Anthropic announcement and same-day coverage from The Register, CNBC, and the Wealth Professional, with templates publicly available in Anthropic's GitHub financial services marketplace.

Go deeper Anthropic finance agents announcement · The Register · TLDR AI

Technology

May 5-6: OpenAI ships GPT-5.5 Instant and open-sources a multi-vendor compute networking protocol

OpenAI replaced GPT-5.3 Instant as the ChatGPT default with GPT-5.5 Instant, citing 52.5% fewer hallucinated claims on high-stakes prompts in medicine, law, and finance, and 37.3% reduction in inaccurate claims on prompts users had previously flagged. The model is now in the API as gpt-5.5-chat-latest and rolls out across consumer plans, with stronger personalization from saved memories, past chats, and connected Gmail and files. Separately, an exclusive in The Deep View detailed OpenAI's Multipath Reliable Connection (MRC) networking protocol — built jointly with Microsoft, AMD, Broadcom, Nvidia, and Intel and released as an open standard through the Open Compute Project. MRC scatters data across hundreds of network paths simultaneously and reroutes failures in microseconds, and it is already in production on the Abilene Stargate site and Microsoft's Fairwater clusters where GPT-5.5 was trained.

So what

For ChatGPT Enterprise customers and any growth company building on the API, the meaningful change is not the headline accuracy gain but the personalization layer — "memory sources" exposes which prior context shaped each response, which materially improves auditability for regulated use cases. Product teams should retest workflows where GPT-5.3 Instant produced flagged errors. The MRC release is the more interesting strategic move: by open-sourcing the protocol with the largest hardware vendors as co-authors, OpenAI is signalling that infrastructure efficiency is a category-wide problem worth coordinating on rather than differentiating with. That posture sits oddly alongside the $4 billion private deployment-services raise reported earlier this week, and the contrast is worth thinking through carefully.

Hype check

The 52.5% hallucination reduction is OpenAI-reported on its own benchmark of "high-stakes prompts"; independent third-party evaluation has not been published, and prior default-model upgrades have shown smaller real-world deltas than launch-day claims.

Confidence: Established fact for the launches; informed perspective on the strategic positioning — based on primary OpenAI publications, The Deep View exclusive interview with OpenAI's networking lead, and corroborating coverage.

Go deeper GPT-5.5 Instant launch · OpenAI MRC paper · The Deep View

Perspectives

BusinessDeveloping

May 5: Vikas Kansal lays out a three-pillar AI paywall framework on Lenny's Newsletter

Building on his Edition #21 piece on why AI economics break the SaaS freemium model, Vikas Kansal — product lead on Google AI subscriptions — published a longer framework on Lenny's Newsletter laying out three monetization levers that align user value with compute cost. Pillar 1: gate usage intensity, structuring tiers (Plus, Pro, Ultra at Google) around volume and context size rather than feature access, since the free tier is already smarter than most users. Pillar 2: gate outcomes, charging not for answers but for completed work, with Intercom's Fin agent at $0.99 per resolution and Sierra cited as the gold standard. Pillar 3: gate the heaviest compute modalities, holding video generation, real-time interactive worlds, and similar compute-extreme features for top tiers as both physical capacity and pricing-ladder signal.

So what

This is the most usable AI pricing framework yet published from inside a frontier lab, and it directly contradicts the seat-based instinct most B2B SaaS teams still default to. CFOs and CROs at growth software companies should bring this to the next pricing review and ask three questions: which features are we accidentally cannibalizing premium tiers with, where could we shift from charging for answers to charging for resolutions, and which compute-heavy capabilities are we giving away that should sit at the top of the ladder. The framework's value is that it finally connects unit economics to upgrade triggers in a way the team can actually implement.

Confidence: Informed perspective — first-person framework from a sitting product leader at Google AI, with named comparable cases at Midjourney, Intercom, and Sierra.

Go deeper Lenny's Newsletter

AI Strategy & Transformation

May 6: "AI-enabled" and "AI-native" are starting to look like cloud-enabled and cloud-native a decade earlier

A piece surfaced today via TLDR Founders argues that the gap between AI-enabled companies (those bolting assistants onto existing processes) and AI-native ones (those rebuilding product, workflow, and org from the floor up) will mirror the cloud-enabled-versus-cloud-native split: identical-looking for a couple of years, then divergent enough that the gap becomes structural. The piece sequences the work as bottlenecks: adoption first — one CTO requires every engineering manager to use Claude Code at the median of their team's usage — then engineering velocity opens up, with a top decile reportedly five times more productive and the median engineer up roughly 20%. After that, the binding constraint becomes product taste: who has the judgment to decide what to build next when the cost of building has collapsed.

So what

The framing is useful precisely because it slows the conversation down. Most growth companies right now are reasonably AI-enabled and assume that is sufficient because the productivity gains feel real. The argument is that those gains are real but bounded, and the companies that rebuild from the floor up will eventually look like a different category of business. The practical implication for a CEO this week: ask your engineering and product leadership where you are on each of the three bottlenecks (adoption, velocity, taste) and whether you are managing for the next one or still working the current one.

Hype check

The 5x and 20% productivity figures are a single CTO's anecdote on his own team; the cloud-enabled-versus-cloud-native analogy is also a hindsight construct that obscures how messy the actual divergence was for any individual company.

Confidence: Informed perspective — single-source essay surfaced via Tier 2 aggregation, well-argued but not independently corroborated.

Go deeper TLDR Founders

Case Studies

Go-to-Market

May 6: OpenAI moves ChatGPT from advertising pilot to self-serve platform, with Shopify embedded

OpenAI launched a self-serve Ads Manager for ChatGPT in U.S. beta, complete with CPC bidding and expanded measurement tools, ending the managed-partnership-only phase that began earlier this year. Advertisers can now create and manage campaigns directly rather than working through agency relationships. In parallel, Shopify rolled out a native integration that lets merchants manage their stores from inside ChatGPT or Claude — adding products, looking up orders, and surfacing insights via prompt rather than dashboard. ChatGPT image generation usage is reportedly up more than 50% in the weeks since the latest model release.

So what

This is a structural shift in how distribution works for any consumer-facing growth company. Advertising inside the assistant is a different placement than advertising on a search results page or a social feed: the user is mid-task, the context is rich, and the click-equivalent action is often a direct purchase or workflow execution rather than a page visit. CMOs at growth companies should run a small test budget through ChatGPT Ads in the next quarter, even if only to gather conversion benchmarks against current channels. The Shopify integration is the more useful case study for B2B teams: it shows the reference pattern for embedding your own product surface inside someone else's assistant without losing your customer relationship.

Confidence: Established fact for the launches; informed perspective on the implications — OpenAI's announcement plus Shopify's product confirmation, with usage figures from OpenAI executive disclosures.

Go deeper ChatGPT for Excel and Sheets · TLDR Founders

Product & Technology

May 5: Subquadratic launches a 12-million-token model that may obsolete RAG scaffolding

Subquadratic emerged from stealth this week with $25 million in seed funding (backers include former SoftBank Vision Fund partner Javier Villamizar and Tinder co-founder Justin Mateen) and shipped SubQ — a 12-million-token-context LLM whose attention mechanism scales linearly rather than quadratically with context length. On long-context retrieval benchmarks, SubQ scored 97% on RULER 128K (versus Opus 4.6's 94%) at $8 per run versus roughly $2,600 on frontier models, and outperformed Opus, GPT-5.4, and Gemini 3.1 Pro on multi-needle retrieval. The team includes researchers from Meta, Google, Oxford, and Cambridge; a 12M-token API and SubQ Code (which loads an entire repository in one pass) are live today, and a 100M-token version is targeted for Q4. Frontier models still lead on reasoning benchmarks like SWE-Bench, where Opus 4.7 holds the top score of 87.6% to SubQ's 81.8%.

So what

If the architecture holds at scale, much of the engineering scaffolding growth companies have invested in to dodge the context-length problem — chunking, embedding, vector databases, recursive prompting — becomes optional rather than mandatory. CTOs should ask their AI engineering teams how much of the current RAG stack exists purely to compensate for context limits, and what the architecture looks like if that compensation is no longer needed. The practical near-term move is smaller: pilot SubQ on workflows where document-grounded retrieval has been frustrating (legal review, multi-document due diligence, large codebase analysis) and measure cost-per-task against the current stack. Cross-session memory remains a separate problem this does not solve.

Hype check

Mamba and earlier sub-quadratic architectures have been pitched as transformer replacements before and have not displaced them; the benchmarks here are credible but the question of whether SubQ scales to frontier-level reasoning at frontier sizes is genuinely open.

Confidence: Watch this — benchmarks are public and API access is live, but durable scaling and frontier-task performance remain to be validated independently.

Go deeper Subquadratic · The Neuron

Resources

Reference

January 2026: Korinek et al. NBER paper sets a roughly six-year clock on automated AI research

Anton Korinek (UVA, Brookings, Anthropic Economic Advisory Council) and three co-authors published an NBER working paper, "When Does Automating AI Research Produce Explosive Growth?", that surfaced this week through The Neuron. The paper models innovation networks under recursive AI improvement and finds that fully automating software R&D, plus modest 5% automation in other sectors, is sufficient to overcome diminishing returns and produce a singularity-style growth take-off within roughly six years under empirically-grounded calibrations. Bottlenecks elsewhere in the economy do not overturn the result if task automation advances quickly enough. The paper is technical but readable, and the author thread on X provides a faster orientation.

So what

This is the most rigorous public modeling exercise yet on whether recursive AI self-improvement can plausibly trigger explosive growth on a defined timeline, and the result is striking enough to deserve careful reading even from skeptics. Use it as the structured reference for board-level conversations on AI strategic horizon — particularly where leadership is dismissive of "intelligence explosion" framing as Silicon Valley narrative. Read it alongside Yann LeCun's countervailing Axios piece this week (which calls 20% job-loss claims "ridiculously stupid") to keep the calibration honest.

Hype check

Korinek serves as a paid consultant to Anthropic's Economic Advisory Council, which is disclosed in the paper but worth weighting; the model's calibration choices are also doing significant work, and reasonable economists will land on quite different timelines from identical inputs.

Confidence: Informed perspective — peer-circulated NBER working paper, not yet peer-reviewed, modeling exercise rather than empirical finding.

Go deeper NBER paper · The Neuron

Sources in this edition: The Neuron, TLDR AI, TLDR Founders, TLDR Marketing, AINews, The Deep View, Lenny's Newsletter, Superhuman AI, The Rundown AI, Bens Bites

← Latest Editions and Search