AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #18

Compute Economics

Thursday, April 30, 2026 · 3 Developments · 1 Perspective · 2 Case Studies · 1 Resource

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

Q1 2026 earnings closed yesterday with a single number that reframes the rest of the conversation: roughly $130 billion in AI capex from four hyperscalers in three months, and Sundar Pichai saying the quiet part out loud — Google Cloud's revenue would have been higher if Google could have built fast enough to meet demand. Several threads from the day fit inside that frame, and together they shift the question executives should be asking from "which model do we use" to "what does the cost curve actually look like for our company."

The throughline today is economic gravity. The capacity, the chips, the cash burn, and the pricing models are all bending in the same direction. The vendor decisions and unit economics most growth companies set even six months ago deserve another look this quarter.

Developments

Business

April 29: Big Tech Q1 capex hits $130B in a single quarter, with capacity already sold out

Microsoft, Google, Meta, and Amazon all reported Q1 2026 earnings on Wednesday, and the headlines converged on one observation: AI is selling faster than the hyperscalers can build for it. Microsoft's AI business hit a $37B annual run rate (up 123% year-over-year), with M365 Copilot at 20M paid enterprise seats and Accenture alone signing 740,000 of them. Google Cloud grew 63% to $20B, but Sundar Pichai said cloud revenue would have been higher had Google built fast enough; the cloud backlog doubled in one quarter to $462B. AWS grew 28% (its fastest in 15 quarters), and Amazon's custom Trainium and Graviton chip business passed a $20B run rate, with OpenAI committing 2 GW and Anthropic up to 5 GW of Trainium capacity. Meta raised 2026 capex guidance to $125–145B, citing memory pricing and additional data-center costs.

So what

Cloud growth rates used to tell you who was winning; now they tell you who has the most concrete poured. For growth-company CEOs, the practical implications are concrete. First, capacity is now a real risk in vendor selection: ask explicitly whether a vendor can guarantee inference capacity at your forecast volumes, and whether their primary cloud has waitlists. Second, expect AI line-item costs to continue rising through 2026 as memory and component pricing eats into hyperscaler margins; build a 20–30% buffer into FY26 AI budgets. Third, the hyperscaler that locks you in on custom silicon (Trainium on AWS, TPU on Google) gets compounding margin advantage. Take a hard look at your cloud commitments before the next renewal.

Hype check

The "demand exceeds supply" narrative is convenient cover for raising capex guidance, and Amazon's trailing-twelve-month free cash flow fell 95% from $25.9B to $1.2B largely because of this spending; if AI revenue does not catch up to the chip bill, the rest of Big Tech's cash flow eventually starts to look the same.

Confidence: Established fact — revenue figures and capex guidance are from the companies' own Q1 earnings releases and 10-Q filings.

Go deeper The Neuron deep dive · Fortune analysis

Competition

April 29: Google to sell TPUs to external customers for installation in their own data centers

Alphabet confirmed it will sell its custom Tensor Processing Units to select external customers, who will install the chips in their own data centers rather than rent capacity through Google Cloud. The company recently announced two new TPU generations targeted at training and inference, and has existing chip-supply deals in place with Anthropic and Meta. Until now, the TPU advantage was monetized through Google Cloud margins; selling TPUs as merchant silicon is a structural change that puts Google in direct competition with Nvidia in the chip market itself, not just in cloud bookings.

So what

For most growth companies the direct buying decision will not come up — this is enterprise data-center silicon. The relevant signal is downstream: a credible second source on AI accelerators changes Nvidia's pricing power, which feeds through to inference pricing within roughly 12–18 months. Add Google's announcement to AWS's Trainium momentum and the merchant-silicon market starts to look like the CPU market did when AMD became credible against Intel. The practical implication for product and infrastructure leaders is that locking into one inference vendor on long-term contracts right now carries more downside than usual; favor commitment levels that match real near-term demand and preserve optionality for the H1 2027 wave of price competition.

Confidence: Established fact — Alphabet has confirmed the program, though the customer list and commercial terms remain undisclosed.

Go deeper Yahoo Finance · TLDR AI

BusinessDeveloping

April 29: OpenAI has effectively abandoned first-party Stargate data centers, will lease capacity instead

Building on yesterday's coverage of OpenAI's Bedrock partnership with AWS, the Financial Times reported on April 29 that OpenAI has in practice abandoned the original Stargate joint venture with Oracle and SoftBank, with one source saying the company has "sidelined first-party data centres." OpenAI itself now characterizes Stargate as an "umbrella term." The company has paused projects in the U.K. and Norway and is leasing capacity from Microsoft and others rather than owning hardware directly. Analysts attribute the pivot to cash discipline ahead of a potential IPO, with some estimating OpenAI could run short of cash by mid-2027 if revenue does not catch up to its compute commitments.

So what

Read this together with yesterday's OpenAI/AWS Bedrock item and the new Anthropic-on-all-four-hyperscalers framing: the foundation labs are decisively becoming software companies running on rented infrastructure, while the hyperscalers absorb the capex risk. For growth companies, the implication is that vendor stability matters more than vendor scale. A foundation lab that cannot finance its own data centers is fine as a partner if it has stable cloud relationships, but it does change the durability profile of multi-year commitments. When negotiating renewals, anchor terms on the underlying cloud provider's commitments rather than the lab's roadmap promises, and treat any "we will own that capacity" claim as worth verifying.

Confidence: Informed perspective — the Financial Times reporting cites multiple sources but is partly anonymous, and OpenAI has not formally confirmed the broad characterization.

Go deeper Tom's Hardware · TLDR AI

Perspectives

Customer Behavior

April 29: Stripe's Emily Glassberg Sands names what is changing — compute is the new cash, and pricing follows

In an extended conversation on Every's "AI & I" podcast published April 29, Stripe's head of data and AI Emily Glassberg Sands made several connected observations from the company's vantage point on tens of thousands of AI-native businesses. Fraud has expanded from a transaction problem to a full-funnel problem, with attackers now stealing tokens from free trials, credits, and unpaid compute bills. AI businesses are scaling revenue faster than the top SaaS cohort ever did, but at materially higher gross-margin pressure because compute costs are real and variable. Outcome-based billing is replacing seat-based pricing in this generation of products. Stripe also announced Link, a consumer wallet purpose-built for delegated agent purchases, signaling that agents-as-buyers is now infrastructure work rather than thought experiment.

So what

Stripe sees the underlying ledgers, so when its data leader says outcome-based pricing is taking over and compute theft is a real fraud category, that is a forward indicator. Three concrete actions follow for growth-company executives. First, audit your own pricing model: if you launched on per-seat and your usage patterns are now agent-driven, your customers are silently overpaying or underpaying, and either case bleeds revenue or trust. Second, ask whether your fraud and abuse controls cover free-trial and credit-tier compute, not just payment fraud; the attack surface has moved upstream. Third, if your roadmap touches B2C, evaluate whether you need to support agent-initiated purchases within twelve months rather than twenty-four. The infrastructure to do it now exists.

Confidence: Informed perspective — Stripe's view is uniquely well-positioned, but Glassberg Sands speaks from one company's data and the macro generalizations should be tested against your own metrics.

Go deeper Every: Compute Is the New Cash · Full transcript

Case Studies

Product & Technology

April 28: Amazon Quick releases as the personal-agent platform integrated with M365, Google Workspace, Salesforce, and Zoom

On April 28, AWS released a desktop app for Amazon Quick, its agentic work assistant, with native integration into the four most widely used business tools and enterprise-grade security and privacy controls already validated by Southwest Airlines, BMW, and the NFL. Quick keeps deep memory across sessions, runs proactive in-background intelligence (flagging unanswered messages, deals needing updates, approvals), takes actions ranging from drafting Slack replies to updating Jira tickets, and creates files including dashboards, presentations, and spreadsheets. AWS also expanded Amazon Connect with prebuilt enterprise-ready agents for hiring, supply chain, healthcare, and customer service. The free version requires no AWS account; sign-in works via Google, Apple, Amazon, or GitHub.

So what

Amazon Quick is positioned for the executive and operations workflow that has been hardest to crack: cross-app, identity-aware, with the audit and compliance posture that makes IT comfortable. Two takeaways for growth-company leaders. First, this is the new minimum bar for "personal AI in the enterprise" — if you are evaluating an agentic productivity tool that does not have native integrations into all four pillars (Microsoft 365, Google Workspace, Salesforce, Zoom), the gap is now visible. Second, AWS's footprint and trust profile in regulated industries means Quick will likely become a default for portfolio companies under enterprise IT control, well before any startup challenger can match the security posture. Run a controlled pilot with one functional team in the next 30 days and benchmark against your current stack.

Confidence: Established fact — the launch is announced and customer references are named, though independent enterprise reviews of the new desktop version are not yet available.

Go deeper Amazon Quick · The Deep View analysis

Go-to-Market

April 29: Perplexity Computer reports $2.8B in labor-equivalent work since launch and pivots into enterprise

At its Ask NYC event on April 29, Perplexity announced that its Computer agent has performed more than $2.8B in labor-equivalent work since launch, including $1.6M of internal Slack-only work in the first four weeks before public release. The company expanded access from Max-tier to Pro subscribers on Mac, added Microsoft Teams availability, launched Computer in Excel as a beta side panel, and shipped over 50 prebuilt enterprise workflow templates with Databricks and Snowflake connectors. A new Computer for Financial Services tier supports bring-your-own-license connectors for Carbon Arc, Daloopa, Morningstar, and PitchBook. A 1Password integration lets the agent take actions in password-protected tools without exposing credentials to the underlying model.

So what

Perplexity's pivot from consumer search to agent-as-employee is the cleanest example yet of what GTM motion looks like when an AI product replaces hours rather than seats: the metric that gets quoted on stage is dollars of labor-equivalent work, not active users. Two practical implications. First, "labor-equivalent work" is becoming the credible value-narration unit for AI products in enterprise sales conversations; if your product can plausibly be measured this way, build the calculation now and use it in pricing. Second, the financial-services vertical play (Carbon Arc, Morningstar, PitchBook connectors) shows how the next leg of agent product growth will come from owning licensed data integrations that competitors cannot replicate. If you sell to verticals with proprietary data sources, locking in those integrations is now a strategic moat rather than an integration backlog item.

Hype check

The "$2.8B in labor-equivalent work" figure is Perplexity's own and is not independently audited; the methodology for converting agent actions into dollar-equivalent labor is unspecified, and similar self-reported "value created" numbers from AI vendors have historically not survived contact with finance teams.

Confidence: Watch this — the product launches and customer-tier expansion are confirmed; the value claim should be treated as marketing until validated by customer references.

Go deeper The Deep View · Perplexity

Resources

Tools

A 90-second multi-model cross-check that catches what either Claude or ChatGPT alone misses

The Neuron's April 29 skill-of-the-day puts a useful frame on something many executives have stumbled into informally: different frontier models make different mistakes, and they are surprisingly good at catching each other's. Microsoft's decision to make M365 Copilot multi-model (auto-routing between OpenAI's GPT and Anthropic's Claude) is the enterprise version of the same idea. The manual workflow takes about 90 seconds and is reproducible without any routing infrastructure: run a meaningful task (a draft, a code review, an analysis) through ChatGPT and Claude in parallel, then paste both responses into a third model with a strict-reviewer judging prompt asking it to identify factual errors, reasoning gaps, and what each response gets right that the other misses. The Neuron published the exact prompt; copy and reuse.

So what

For executives who already pay for both Claude and ChatGPT seats (most by now), this is essentially free quality uplift for high-stakes outputs: board memos, customer escalations, M&A diligence reads, code reviews. Add it to your team's standard operating procedure for any document or analysis where being wrong costs more than 90 seconds saves. For technical leaders, the same logic argues for adding cross-model verification as a checkpoint in agent workflows handling production decisions, not just as a manual habit.

Confidence: Informed perspective — the underlying observation that models have different blind spots is well-established; the specific 90-second workflow is The Neuron's framing.

Go deeper The Neuron skill of the day

Sources in this edition: The Neuron, TLDR AI, Every, The Deep View, Financial Times (via Tom's Hardware), Fortune, CNBC, Semafor

← Latest Editions and Search