Print, Not Pilot

Two earnings prints do not yet make a structural inflection. NVIDIA's $75.2 billion data center revenue includes hyperscaler capex commitments that are years of forward visibility — if any of Microsoft (committed $190 billion CY2026), Meta ($125-145 billion), or Amazon (~$200 billion) cuts its 2027 capex by even 15 percent on a ROI reassessment, NVIDIA's Q2 FY28 starts shipping into a structurally weaker order book. Workday's $500 million agentic ARR is the same materiality category as Salesforce Agentforce ARR ($800 million at 169 percent YoY) and ServiceNow's Now Assist disclosures — incumbents are pricing per-action AI on top of per-seat contracts, but the renewal-cycle test arrives only when net-new ACV needs to absorb seat compression rather than expansion. The agentic-ARR-is-real story is now well-supported; the agentic-ARR-is-net-additive story is not. Goldman Sachs's May 2026 productivity note — that there is no meaningful relationship between AI and productivity at the economy-wide level — is the disconfirming evidence on whether the buyers of this $500 million ARR get measurable returns.

Named-customer disclosures at vendor conferences are a vendor-vetted dataset, not an independent one. KPMG's $120 million contract-leakage target is exactly that — a target, not a realised number; Ericsson's 90,000 hours saved by 85,000 employees averages out to roughly one hour per employee, a productivity claim small enough that it could be measurement bias and large enough that SAP wants to put it on the keynote slide. The harder question is whether SAP's Autonomous Enterprise is a defensible product moat or a price-protection wrapper — Workday and Salesforce already have analogous offerings at comparable scale (Agentforce $800 million ARR at 169 percent YoY; Workday agentic approaching $500 million), and the Microsoft Azure deepening means SAP is now structurally dependent on a hyperscaler that is simultaneously a competitor for the data-and-agent layer. The saaspocalypse pattern is now bifurcated: incumbent ERP and CRM platforms catching the wave at named-customer scale, while the underlying public software multiples have collapsed below the S&P 500 P/E for the first time in the modern era.

Benchmark Theater applies here: Terminal-Bench 2.1 and GDPval are benchmarks where Google has direct authorship or co-authorship influence, and the 76.2 percent score is Google-reported on a leaderboard where Google publishes its own evaluations — treat as directionally correct, not independent validation. The $100 AI Ultra tier (down from $250) is the giveaway: Google is competing on price because OpenAI's GPT-5.5 Instant (now ChatGPT default with High Capability cybersecurity / biology classification) and Anthropic's Claude Mythos (restricted to Project Glasswing partners at 93.9 percent SWE-bench Verified) hold the absolute capability ceiling, not Gemini 3.5 Flash. The Cloud backlog story (Sundar Pichai noting customers exceeded initial spend commitments by 45 percent) is structurally strong but DOJ antitrust risk and ChatGPT search substitution remain real — the bear case isn't dead, it has just been deferred.