AI Beats and Backlogs: A Tale of Four Companies

$460 billion — Google Cloud's signed backlog at the end of Q1 2026, after it nearly doubled in a single quarter. That's more than two times Google Cloud's trailing-twelve-month revenue. It's the line in tonight's earnings that turned all four hyperscaler reports from a beat into a verdict. Three years of debate about whether AI demand was real ended with one number. Now Wall Street is asking the next question — whose AI capex is showing up as AI revenue, and whose is still a roadmap. Same beat. Three different verdicts.

Stop reporting "AI capex" as one number on your next investor update. Every audience for a 2026 strategy update — board, investors, customers — wants to know what backlog is backing the spend. Google walked into earnings with a $460 billion signed backlog against a $185 billion capex guide. Meta walked in with $145 billion of guided spend and no equivalent backlog metric whatsoever — and got punished accordingly, despite faster top-line growth. The dollar amount is no longer the headline. The contracted-demand-to-capex ratio is. Build the slide before someone on your board asks why you don't have it.

Audit your CPM-to-CPA spread on every ad platform you buy this quarter. Google Search queries hit an all-time high tonight because agents are now using Google more than humans do. Instagram impressions, Facebook impressions, every ad-funded surface — same dynamic, growing fast. The platforms bill you on CPM. The reason you're in business is CPA. The wedge between those two numbers is the agent tax, and it lands in your P&L, not theirs. Pull conversion data by user-agent. Anything that's flat or rising on impression cost while flat or falling on conversion volume is your agent tax showing up. Most marketing teams don't have a line for it yet. The ones that build that line in 2026 will renegotiate inventory in 2027.

Run Lemkin's audit on every line item in your software stack. Is this critical to my AI agents being successful at their jobs? SaaStr paid Salesforce 83% more last year because the agents query it 100x more than the humans ever did. They're about to cancel Notion because the agents don't care. The yes pile gets more spend, more access, more permissions. The no pile is dead weight you cut now or watch stealth-churn in the next four quarters. The renewals won't tell you the truth in time. Run the audit yourself today.

Keeping up with AI is hard. We know — we do this daily. If any of our action items strike a chord, or you simply want an Outsider perspective on how you can improve your business with AI, we'd love to hear from you.

Outsider Labs Advisory →

Wednesday, April 29

1. The bubble argument died at 4:01 PM. The bifurcation argument started at 4:02. All four hyperscalers beat AI quarter — Google Cloud +63%, AWS +28% (fastest in 15 quarters), Azure +40%, Meta revenue +33%. Combined Big 4 revenue: $430.6 billion. Combined Q1 capex: $112 billion. And yet by 5:30, Alphabet was up 7% after-hours while Microsoft and Meta were both red — Meta down nearly 7% on a beat-and-raise. Same night, three different verdicts. The market just changed the metric from capex to backlog, and only one company had a $460 billion line item to point at. (Full analysis below.)

2. Microsoft is now being punished for the sins of OpenAI. Microsoft's $37 billion AI run rate (+123% YoY) is real and not OpenAI-dependent. The market is discounting it anyway. The 27% stake in OpenAI used to be the most leveraged passive position in AI; tonight it's the most visible drag. WSJ reported Friar is worried OpenAI can't fund its compute commitments — $600 billion across Azure, AWS, and Oracle against a $30 billion revenue target with a $25 billion burn. Twenty-four hours after losing exclusivity, OpenAI walked into Amazon's Bedrock with an exclusive Managed Agents runtime. Sam Altman appeared at the launch by recorded video because he was in Oakland for the trial. (Full analysis below.)

3. A Claude Opus 4.6 agent wiped a startup's entire production database in nine seconds — on the same night Claude went down. PocketOS asked the agent to fix a credential mismatch in staging. Instead it ran a deletion command via the infrastructure provider and wiped both the production database and all backups. Post-mortem from the agent itself: "I guessed instead of verifying." Claude separately had a major outage hours later, on the same evening Big Tech reported record AI revenue. The infrastructure-reliability gap and the agent-governance gap, in two stories, in one news cycle. If you're running agents in production this quarter, your "least privilege" policy needs to ship before Friday.

4. Figure is now making humanoid robots at the same cadence Toyota makes Camrys. 350 Figure 03 robots delivered. Production rate at one robot per hour — a 24x throughput improvement in 120 days. End-of-line first-pass yield north of 80%. Sam Altman, separately, called general-purpose robotics "the new chess piece" the U.S. needs to rival China and confirmed OpenAI is building a dedicated robotics team. Sony's robot just beat professional table tennis players. GM is rolling Gemini to four million vehicles via OTA. The robotics narrative just crossed from "demo videos" to "fleet operations" — and nobody on the earnings calls quantified what that does to industrial capex.

5. David Silver just raised $1.1 billion seed at a $5.1 billion valuation on the bet that human training data is running out. The man who built AlphaGo and AlphaZero is the one telling Sequoia, Lightspeed, Google, and NVIDIA that the LLM era's training corpus has a floor. Ineffable Intelligence is built on RL-from-experience — agents that learn from interaction with the world rather than scraped text. If Silver is right, every "scaling laws are all you need" piece written between 2022 and 2025 just got challenged by the guy who shipped reinforcement learning at superhuman scale before the rest of us knew what a transformer was. The labs that own simulators and physical environments — xAI, Tesla, Figure — get a structural advantage. The text-scrape labs get a problem.

The Metric Just Changed

For the first time since October 2020, all four hyperscalers reported earnings the same night. Alphabet, Microsoft, Meta, Amazon. Every one of them beat. Every one of them credited AI. Combined revenue: $430.6 billion. Combined Q1 capex: $112 billion, on pace to crack $600 billion for the year. Google Cloud grew 63% to $20 billion — fastest in its history. AWS +28% (fastest in 15 quarters). Azure +40%. Microsoft's AI business hit a $37 billion annual run rate, up 123% YoY. Meta revenue grew 33% — the fastest top-line print in the four. The bubble argument died at the closing bell.

And yet, by 5:30, Alphabet was up 7% after-hours, while Microsoft and Meta were both red — Meta down nearly 7% on a beat-and-raise. Same night. Same earnings calendar. Same AI thesis. Three opposite reactions. That's the story.

The cleanest read on the split: Wall Street just changed the metric. For two years the question was whether the hyperscalers would commit enough capex to win the AI cycle. They did — $112 billion in one quarter, with Alphabet selling a 100-year century bond (the first by a tech company since Motorola in 1997) to help fund it. The capex commitment is now done. The market priced it. The new question is what's the contracted demand backing it — and the answer separates the four companies into completely different stories.

Google has the receipts. $460 billion in signed Cloud backlog, doubling QoQ. 330 customers each processing more than a trillion tokens. Customers outpacing initial AI commitments by 45%. Pichai literally said cloud revenue would have been higher if Google could meet the demand — they're compute constrained. AWS has its version: $364 billion of backlog plus $100 billion newly signed, plus an exclusive Bedrock Managed Agents runtime that OpenAI brought to Amazon within twenty-four hours of Microsoft losing exclusivity. Microsoft has the run-rate version, $37 billion at +123%. It's real. It's just shadowed by the 27% of OpenAI attached to it — and that drag is now visible to every procurement team running the Christie Brinkley test on Sam Altman.

Meta has none of those. What it does have is consumer impressions, a 33% revenue print, 3.5 billion daily users, and a capex guide raised to $125-145 billion. What it doesn't have is a backlog metric — or a Cloud business — or an enterprise procurement queue. And here's the part nobody on the call asked about: a growing share of the impressions Meta sells on a CPM basis are now being delivered to agents instead of humans. Every ChatGPT browsing session that crawls Instagram for product reference. Every Manus agent reading Facebook Marketplace. The advertiser pays the CPM. The advertiser lives or dies on CPA. Agents drive the impression count up while contributing nothing to acquisition. The platform's revenue rises while the customer's unit economics get worse. That setup always reprices.

Mark Zuckerberg keeps describing this as "personal superintelligence to billions of people." It is not curing cancer. It is a higher-resolution ad targeting engine — doing a better job serving up Sydney Sweeney pictures and French Bulldog videos to a feed that's increasingly being scrolled by bots. That's a real business — Google did it for twenty years and printed money. The structural difference is that Google has a path to monetize the agent layer itself (Gemini Enterprise, indexing licenses, sponsored agent answers). Meta does not. Search revenue hit an all-time high tonight precisely because agents query Google more than humans ever did. Meta's ad auction has no equivalent. Same problem, two different houses, two different sets of keys. Google has the keys. Meta is still figuring out which door to walk through.

The forwardable line: all four hyperscalers won the AI capex argument tonight, and only one of them won the AI revenue argument. That spread is going to compound in 2027 the way mobile-search spreads compounded in 2012. Pick the side of the receipts.

At COAI today: Full analysis at getcoai.com — the bifurcation, the Christie Brinkley problem at Microsoft, and the agent tax hiding inside Meta's beat.

— Harry and Anthony

Sources: