The NextWave Signal
Week of 3–9 May 2026 | Issue #5
This week on the NextWave Signal
The Inference Price Collapse Hasn’t Fixed Enterprise AI Economics — The cost of querying a frontier model has fallen 280-fold since late 2022. Token prices dropped another 80% in 2025 alone. None of it matters as much as it should: token costs now represent under 20% of enterprise AI total cost of ownership. The real drag is data engineering overhead (25–40%), model maintenance (15–30%), and legacy system integration (2–3x premium). Eighty-five percent of organisations still misestimate AI project costs by more than 10%. Read →
Export Controls Built a More Efficient Chinese AI Ecosystem — US chip restrictions blocked China’s access to Nvidia H100s. DeepSeek’s response — building R1 on H800s using mixture-of-experts and grouped query attention — was not a workaround. It was an efficiency research programme accelerated by compute constraint. CSIS frames the compound risk plainly: when China also develops its own leading-edge chips, the resulting ecosystem will have both the capacity and the algorithms. Hardware-only export controls have predictable limits. Read →
Enterprise AI Pricing Has Split in Two — ServiceNow bundled its AI capabilities into existing contracts (Now Assist approaching $500M ACV). Salesforce raised prices 6% and reported a revenue beat. Both strategies have historical precedent — and both result in the same enterprise outcome: $30–50/user/month in AI software costs becoming a fixed operating line before any purpose-built tooling. Read →
IBM’s Confluent Acquisition Names the Real Bottleneck — IBM paid to solve data freshness, not model quality. The $4.7B deal signals that the enterprise AI constraint in 2026 is stale data reaching models at decision time — not the models themselves. Every major enterprise platform stack (SAP, Salesforce, Oracle) faces the same integration debt. Read →
Three Federal AI Copyright Rulings, Three Different Outcomes — Thomson Reuters won on market substitution. Alsup found Anthropic’s training transformative but piracy separately infringing ($1.5B settlement). Chhabria found Meta’s piracy fair use but flagged a market-dilution theory that future plaintiffs now have time to build. Enterprise vendor indemnities cover output infringement only — not the gaps these rulings exposed. Read →
GLP-1’s Constraint Is Manufacturing, Not Science — Eli Lilly’s $6B Alabama plant and Amgen’s $1.4B Ohio expansion are supply chain engineering projects, not drug development. Oral GLP-1 formulations could reshape the economics before the injectable capacity comes online. Read →
Signals worth watching
- Big Tech Q1 2026 earnings — AI capex acceleration confirmed — Microsoft, Alphabet, and Meta all reported in the final days of April. The headline across all three: AI infrastructure capex is accelerating, not moderating, into H2 2026. Cloud AI revenue growth is outpacing overall cloud growth at all three companies. The buildout is structural, not cyclical — consistent with what TSMC’s 66% gross margins already implied. (Microsoft IR)
- Colorado AI Act — 52 days to June 30 — The Colorado Attorney General’s enforcement guidance is now live. The Act applies to developers and deployers of high-risk AI systems affecting Coloradans. Organisations without compliance documentation in place are now in a structurally difficult position — the 72-day assessment cycle required for full compliance already requires more time than remains. (Colorado AG Office)
- EU AI Act harmonised standards remain incomplete with August deadline approaching — The August 2026 deadline for high-risk AI system compliance is 13 weeks away. The European Commission’s Digital Omnibus proposal adds procedural uncertainty without moving the compliance date. Organisations relying on the standards to be completed before starting implementation are waiting for something that will not arrive in time. (European Commission)
Stat of the week
85% — the share of organisations that misestimate AI project costs by more than 10% before implementation begins, per Xenoss analysis of enterprise AI total cost of ownership. The primary driver is underestimating data engineering overhead, not the API bill.
Xenoss Enterprise AI TCO Analysis
What to watch next week
- Google I/O (May 14) — The annual developer conference is the most significant AI product announcement window outside of OpenAI’s own events. Watch for Gemini 2.5 deployment timelines, TPU availability, and any shift in Google’s open vs. closed model strategy after Gemma 4’s Apache 2.0 release.
- Colorado AI Act enforcement clock — 45 days to June 30 as of next Friday. Watch for enterprise legal teams beginning to publish their compliance status publicly, which will accelerate pressure on deployers who haven’t started.
- OpenAI IPO developments — The $852B valuation and enterprise API repricing implications remain unresolved. Any regulatory comment period developments or S-1 preview signals will matter to enterprise buyers evaluating long-term API dependency.
— The NextWave Insight editorial team
