Week 18 · April 27 – May 2, 2026

high signal

The Earnings Print

Big Tech told the market it would spend roughly $700 billion on AI infrastructure in 2026 — nearly double 2025 — and the market sold most of them. Microsoft, Alphabet, Amazon, and Meta all beat expectations on the print and most fell anyway, with the punch landing on the capex line, not the revenue line. ServiceNow beat every metric and raised its 2026 AI revenue forecast 50 percent — and lost 18 percent in a single day, the largest one-day drop on record, as management disclosed half of net new business now comes from non-seat models. GPT-5.5 shipped April 23 as the first fully retrained OpenAI base since GPT-4.5, and both it and Claude Opus 4.7 cleared the 72.4 percent human-expert baseline on OSWorld-Verified for autonomous computer use. Q1 closed with 78,557 tech layoffs and 47.9 percent explicitly attributed to AI — versus under 8 percent a year ago. Cursor's $2 billion round at a $50 billion valuation hit oversubscribed within days, doubling the company's mark from six months ago. NIST's first independent evaluation of DeepSeek V4 Pro placed it eight months behind the US frontier — wider than DeepSeek's own self-reported gap. The capex bill arrived. The seat economy started repricing. The labor data stopped being theoretical.

5 predictions updated4 milestones20 companies refreshed

Key Developments

Big Tech committed roughly $700 billion to 2026 AI capex — and most of the stocks fell on the print

Microsoft, Alphabet, Amazon, and Meta all reported Q1 2026 in the same week and the combined 2026 capex commitment landed near $700 billion against roughly $380 billion last year — a near-doubling at the largest scale ever attempted. Microsoft guided $190 billion, Alphabet $180-190 billion, Amazon $200 billion, Meta $125-145 billion. The fundamentals beat: Azure reaccelerated to +40 percent year over year, Microsoft's AI business hit a $37 billion run rate (+123 percent), Copilot paid commercial seats reached 20 million from 15 million in January. Google Cloud crossed $20 billion quarterly revenue for the first time at +63 percent, with $460 billion in backlog. AWS grew 28 percent — its fastest in 15 quarters — on a $15 billion AI ARR. Meta's revenue ran +33 percent on AI-improved ad targeting. Microsoft and Meta both fell on the print. Amazon's free cash flow collapsed from $25.9 billion trailing-twelve-month a year ago to $1.2 billion this quarter as $59 billion of incremental property and equipment hit the balance sheet. The market is now pricing AI as cash burn first, revenue second.

Confirms P-038 ↗ Approximately 1% of total US economic growth in 20...Confirms P-021 ↗ 40% of enterprise applications will feature task-s...Confirms ↗

Counterpoint

Treat the $700 billion as a guidance ceiling, not a check already cut. Microsoft's $190 billion and Amazon's $200 billion are full-year intent figures issued in late April with three quarters left to walk back; Meta's $125-145 billion has a $20 billion swing baked in. Roughly $46 billion of the combined Q1 reported profit at Alphabet ($28.7 billion) and Amazon ($16.8 billion) was unrealized mark-to-market gains on their Anthropic equity stakes — accounting profit, not cash inflow, and reversible if Anthropic's secondary marks compress. Microsoft and Meta selling off on a beat is not the AI thesis breaking; it is investors discounting the ROI horizon, which they have done at every major capex cycle (cloud 2014-15, mobile 2010-11) before the spending paid off.

Source →

ServiceNow beat every Q1 metric and raised AI guidance 50 percent — then lost $30 billion of market cap in one day

ServiceNow reported Q1 revenue of $3.77 billion (+22 percent), beat operating margin (32 percent vs 31.5 percent guided), and raised its 2026 AI revenue forecast about 50 percent to roughly $1.5 billion. The stock opened down 18 percent — its worst single-day drop on record, evaporating $30 billion of market cap. The disclosure that did the damage was on the call: management said about half of net new business now arrives through non-seat models — token consumption, per-action billing, agent connectors. Salesforce dropped 9 percent in sympathy on the same day. The iShares IGV software ETF is now down 21 percent year-to-date and 30 percent from its September 2025 peak; software's forward P/E sits at 22.7x — below the S&P 500 multiple for the first time in the cloud era, after sitting near 84x in 2021. Beating earnings is no longer enough; the unit of value is changing under enterprise software's feet, and the multiple is repricing before the revenue model finishes the transition.

Confirms ↗ Confirms P-021 ↗ 40% of enterprise applications will feature task-s...

Counterpoint

A 22 percent revenue grower is not a melting ice cube. ServiceNow's consumption pivot is what the bulls have been asking software companies to do for two years; the same investors who punished the print would have punished a Q1 that showed seat growth slowing without a consumption replacement. Goldman Sachs published its six-factor AI Impact Framework on April 13 with a specific buy list — institutional capital is sorting winners from losers, not fleeing the sector. Salesforce's Agentforce ARR hit $800 million (+169 percent year over year) and 29,000 enterprise deals as of W17 — the line item bull case is intact even while multiples compress. The honest read: the IGV drawdown is real, the structural repricing of the seat economy is real, and individual SaaS names trading at S&P-500 multiples will look cheap in 12 months if Agentforce-class lines continue compounding through the transition. The Q2 prints on seat counts at Workday, ServiceNow, and Salesforce will settle the argument.

Source →

GPT-5.5 cleared the human-expert baseline on autonomous computer use — and so did Claude Opus 4.7

OpenAI shipped GPT-5.5 on April 23, the first fully retrained base model since GPT-4.5 and not a post-training iteration on the GPT-5 trunk. On OSWorld-Verified — the benchmark that measures whether a model can drive a real desktop, navigate GUIs, and complete multi-step tasks unsupervised — GPT-5.5 scored 78.7 percent against a 72.4 percent human-expert baseline. Claude Opus 4.7 came in at 78.0 percent on the same benchmark, the first time two frontier models cleared the human-expert bar on autonomous computer use in the same release window. GPT-5.5 also took Terminal-Bench 2.0 at 82.7 percent (a 13.3 percentage-point gap over Opus 4.7), GDPval at 84.9 percent across 44 occupations, and FrontierMath Tier 4 at 39.6 percent. JetBrains' January 2026 survey of 10,000 developers landed in late April and put the deployment number alongside the capability number: 90 percent of developers use AI tools at work, Claude Code went from 3 to 18 percent workplace adoption in six months (a 6x ramp), and 51 percent of code committed to GitHub is now AI-assisted.

Confirms P-001 ↗ AI models will handle most aspects of software eng...Confirms P-021 ↗ 40% of enterprise applications will feature task-s...Confirms ↗

Counterpoint

One caveat: OSWorld is a controlled lab benchmark, and GPT-5.5 ran on a harness OpenAI has visibility into — Benchmark Theater again unless an independent third party (METR, NIST, academic re-test) reproduces the 78.7 percent number. A 21 percent failure rate on a curated benchmark also compounds fast in production: chain three OSWorld-class tasks together and end-to-end success drops to 49 percent, chain six and it hits 25 percent. The 51 percent AI-assisted-commit figure from JetBrains conflates AI-suggested completion accepted by a human with fully generated code; Stack Overflow's independent number is closer to 30 percent for the latter. And "exceeds human-expert baseline" sounds civilizational but means something specific — beating an evaluator who is already AI-assisted at task documentation. The gap from benchmark capability to "AI agents do work without human review" — the language of P-021 — remains the entire game.

Source →

What the Evidence Moved

P-001AI models will handle most aspects of software engineering t...

GPT-5.5 (April 23) and Claude Opus 4.7 both cleared the 72.4% human-expert baseline on OSWorld-Verified for autonomous computer use (78.7% and 78.0%) — first time two frontier models cleared an autonomous-operation human bar in the same release window. JetBrains' 10,000-developer survey: 90% use AI tools at work, Claude Code went 3% → 18% workplace adoption in six months, 51% of GitHub commits are AI-assisted. Cursor's $2B oversubscribed round at $50B and $2B ARR confirms the productivity gain is monetizable at AI-native pricing. The capability and adoption sides moved together this week.

73%→78%▲ +5pp

P-002AI will disrupt 50% of entry-level white-collar jobs over 1-...

Q1 2026 cumulative tech layoffs reached 78,557 with 47.9% explicitly AI-attributed in employer statements — versus under 8% a year ago. Oracle 30,000 (12,000 India), Meta 8,000 starting May 20 plus 6,000 cancelled requisitions = 14K reduction. Goldman Sachs net US AI-driven job destruction model: ~16,000/month. Anthropic observed-exposure data: programmers 74.5%, customer service 70.1%, data entry 67.1%. Indian IT entry-level roles down 20-25% per EY citation. The empirical floor under the prediction firmed; the Fed's March FEDS Note still shows aggregate job postings holding, so disruption is compositional rather than headline-unemployment.

62%→65%▲ +3pp

P-020AI-exposed industries will see job losses of ~20,000 per mon...

Same Q1 layoff cluster as P-002, with one specific addition: Goldman's substitution-to-augmentation ratio widened (25K substituted vs 9K augmented per month). Meta's 14K reduction is the cleanest single-name AI-driven structural cut of 2026 to date — explicitly framed as capital reallocation from headcount to AI infrastructure. The structural, not cyclical, framing of the cuts (pod reorganization, role-category renaming) is what the prediction was calibrated against.

60%→63%▲ +3pp

P-02140% of enterprise applications will feature task-specific AI...

GPT-5.5 OSWorld-Verified at 78.7% and Claude Opus 4.7 at 78.0% both clear the 72.4% human-expert baseline for autonomous computer use — the capability layer the prediction tracks. Microsoft Copilot paid commercial seats hit 20M (+33% sequential from January's 15M), validating the seat count required for Gartner's 40% enterprise-app-with-agents threshold. Salesforce Agentforce at $800M ARR + 29,000 enterprise deals (W17). Gartner's own updated data showed 80% of new enterprise apps shipped or updated in Q1 2026 now embed at least one AI agent vs 33% in 2024. The platform layer materialized faster than Gartner originally modeled.

83%→87%▲ +4pp

P-012AI investment boom is a bubble likely to burst by 2026

URGENT — target_date 2026-06. The IGV software ETF correction (-21% YTD, -30% from peak) and ServiceNow's -18% single-day drop are real re-pricing events but do not meet a bubble-burst definition (50%+ NVDA drawdown, H100 prices crashing, broad AI capex retreat). Polymarket assigns 13-17% to a bubble burst by year-end 2026. With ~30 days to target_date and no triggering catalyst visible, p=0.05 holds. P-012 will resolve refuted at 2026-06-30 absent a discrete shock event.

5%→5%▼ 0pp

Company Impact

Caterpillar

Score change

Q1 2026 — revenue $17.42B (+22% YoY), EPS $5.54 vs $4.62 expected, power generation revenue +41% to $2.82B explicitly tied to data center AI demand. Record backlog $63B (+79% YoY); full-year guidance raised to low double-digits vs prior ~7% CAGR. Cat AI Assistant launched in Q1 with voice in-cab + remote diagnostics. AI infrastructure exposure now revenue-confirmed, not projected. are 6→7, score 7.8→8.1 (crosses very_positive threshold).

CNBC CAT Q1 · Bloomberg AI boom · caterpillar.com Cat AI Assistant

Key Developments

Big Tech committed roughly $700 billion to 2026 AI capex — and most of the stocks fell on the print

ServiceNow beat every Q1 metric and raised AI guidance 50 percent — then lost $30 billion of market cap in one day

GPT-5.5 cleared the human-expert baseline on autonomous computer use — and so did Claude Opus 4.7

What the Evidence Moved

Company Impact

Caterpillar

Microsoft

Alphabet / Google

Amazon

Meta Platforms

Apple

Oracle

ServiceNow

Cursor (Anysphere)

OpenAI

Cerebras Systems

Broadcom

Booking Holdings

Citigroup

Costco

Charles Schwab

Cisco Systems

CrowdStrike Holdings

Concentrix

Coca-Cola

Comcast

Colgate-Palmolive

Chevron

Charter Communications

Chegg

Bristol-Myers Squibb

Capital One Financial

C3.ai

CVS Health

Funding Snapshot

Regulatory Calendar

Capability Milestones

Sources