Menlo’s 2026 enterprise AI map: falsify agent-driven spend structure—not “SaaS is dead”

“SaaSpocalypse” assumes AI zero-sum cannibalizes existing software or labor budgets. Menlo’s fourth annual enterprise AI report draws a different structure: deployment spend jumps to ~$177B, over half of incremental dollars from new budgets, with the app layer concentrated where agents work—coding and chat.

Falsifiable read: the point is not “software is dead,” but whether spend truly scales with agent work rather than seats, and whether lab/app shares replicate in independent sources. If cloud filings, procurement panels, and lab reports cannot support structures like 5.4× and 69%/76%, discount “agents rewrote the enterprise AI spend equation.”

Hard facts (Menlo Ventures; Tim Tully, Joff Redfern, Deedy Das, Derek Xiao; 3 Nov 2026; methodology: survey of nearly 400 U.S. enterprise decision-makers responsible for >$4B in AI spend, fielded by Morning Consult 8–26 Sep 2026; market sizing is Menlo bottoms-up estimates):

• Totals: deployment ~$176.8–177B (Key Findings say $177B; body $176.8B) from $33B, 5.4× (vs 3.9× the prior year). Development (building models) ~$102.5B vs deployment/using ~$176.8B—about $1.70 of use-side revenue per $1 into training.
• Lab share: Anthropic+OpenAI+Google ~$122.7B, 69% of deployment dollars (50% last year). Enterprise LLM API: Anthropic 40%→55%, OpenAI 27%→31%, Google 21%→9%; the two leaders 86% of API (67% last year). Open-source: usage 11%→19%, spend 6%→4%.
• Apps: app layer $77.3B; coding $34.0B + chatbots $25.0B = 76%; coding 96% AI-native; Claude Code+Codex ~$24.0B (70% of coding).
• Agents & budgets: on OpenRouter agents ~80% of tokens (~7× human), agent traffic +37× since overtaking humans in February; average enterprise spend 4.7× last year vs 1.9× planned; 55% of AI dollars from net-new budgets. Deployment spend >30% of public-cloud revenue (from ~6%).

Evidence strength: VC annual perspective + commissioned survey + bottoms-up market model, with OpenRouter and other third-party traffic inputs. Strength is structural granularity (labs/apps/open-source/native vs incumbents); limit is that dollar totals and shares are Menlo estimates triangulating survey weights and public disclosures—not a regulator/audited census. Some named companies are Menlo portfolio firms.

Steelman: if agents expand consumption with work rather than seats, budget overruns and spend concentration in verifiable environments (coding, support) are two sides of one mechanism—more testable than “AI replaces Salesforce.”

Strongest steelman: filtering categories by “environments agents can act in + learnable context + verifiable outcomes” explains why coding/chat sprint while HR/marketing lag—a portable industrial hypothesis, not just a leaderboard.

Steelman risks: OpenRouter skews developers/startups; API shares depend on self-reported production mix; the ~$177B level is taxonomy-sensitive (prior years reclassified); portfolio conflicts are disclosed via asterisks and should be discounted by readers.

For buyers: extend beyond model brands to agent infrastructure (sandboxes/search/browser adoption cited at 40%/30%/14%) and verifiable workflow design. For legacy SaaS: the report finds limited evidence of wholesale SoR replacement—more threat is growth dilution than overnight swap. For open-source: usage up, dollar share still thin; price wars and cache-hit mixes keep pressing average $/token down.

Do not read Menlo’s map as a “SaaS is dead” verdict. Falsify whether agents keep raising deployment-per-seat, whether lab dollars stay concentrated, and whether non-coding verticals can install act/verify conditions. If independent cloud bills, procurement panels, and lab filings align, the story hardens; if not, dial it back to one VC’s high-resolution hypothesis.

【No.2】