The $650B Build-It-and-They-Will-Come Moment in AI

Source: Gemini

Hyperscalers make big bets on supply — and big waves driving demand

According to Bridgewater, Big Tech hyperscalers plan a massive $650 billion AI infrastructure investment in 2026. Yet Menlo Ventures reports 2025 enterprise AI spending hit just $37 billion. That gap suggests suppliers expect a major surge in AI service spending. Or it’s just FOMO. Or both. Reuters Menlo Ventures

Source: Ben Evans

To spur demand, Microsoft announced Copilot Cowork — an agentic application built with Anthropic’s help to automate non-engineering workflows. The agent executes plans in the background using Microsoft 365 data from Outlook, Teams, and SharePoint, while operating within enterprise security, privacy, and governance boundaries. link

The demand equation extends beyond direct enterprise implementations. Alphabet reported Google Search revenue grew 17% year-over-year to $63 billion, driven by AI Mode queries that have doubled since launch and run three times longer than conventional searches — creating new ad inventory as AI reshapes how people seek information. link

Security moves to center stage

That wave of new hyperscaler AI tools has already rattled markets. Announcements of Claude Code Security, Google CodeMender, and OpenAI Aardvark triggered a cybersecurity stock selloff — what Forrester called a “SaaS-pocalypse.” The anxiety isn’t unfounded: Google reports only 27% of security leaders feel equipped to secure applications against autonomous, agent-driven coding. Forrester report

Source: Forrester Research

The OpenClaw craze illustrates the stakes. The unleashed local agent sold out Mac Minis and racked up massive downloads — but also wiped a security researcher’s inbox against her explicit instructions. OpenAI has since hired founder Peter Steinberger to build next-gen personal agents, while transitioning OpenClaw to a foundation-supported open-source project. X Cisco

Source: Summer Yue/Simon Willison

The tide seems to be rising

The ROI picture remains murky. An Andreessen Horowitz survey finds enterprise AI returns mostly hovering between 1x (breakeven) and 1.5x — solid if not spectacular. link

Source: Andreessen Horowitz

An NBER study of 6,000 global executives confirms the pattern: little AI impact to date, though firms predict three-year gains in productivity (1.4%) and output (0.8%), paired with a 0.7% employment drop. Employees are more optimistic, expecting a 0.5% employment boost. Notably, CEOs use AI more than the rest of the C-suite. link

Source: NBER

To drive results, Google developed a KPI framework for production AI agents spanning operational reliability, adoption, and business value. Gartner adds five board-ready metrics — sales conversion rates, labor costs, time to value, collection efficiency, and employee NPS — to make the ROI case stick. Google Gartner

Wrestling with productivity

Harvard Business Review reports that AI is intensifying work, not reducing it. Workers using AI tackle a broader range of tasks with less downtime and more multitasking. Getting this balance right is critical for both employees and the employers expecting efficiency gains. link

Agency economic models are under pressure to evolve. S4 Capital’s Monks expects 25% of its revenue to be subscription-based by year-end — the logic being that as AI improves productivity, clients receive more output for the same fee rather than fewer billable hours. link

Working on workforce

Bain argues unlocking AI value requires simultaneously redesigning workflows and modernizing workforces. As agents deploy, work shifts from doing to planning and reviewing. The firm also notes that AI failures are often design flaws, not model errors — and perceived AI disruption risk is rising across all industries. link link

Source: Bain

Block is cutting 40% of its workforce and attributing it to AI. Wary observers suspect it’s mostly right-sizing after the company doubled its headcount between 2020 and 2024. Both can be true. link

Swings and roundabouts

Enterprise vendor trust is being tested. Anthropic refused Pentagon demands for unrestricted use of Claude and was blacklisted by the Trump administration. OpenAI signed a similar deal the next day, triggering 1.5 million #QuitGPT cancellations in 48 hours. For enterprise buyers, this raises real questions about vendor alignment. Meanwhile, Anthropic hit a $19B annualized run rate and Claude jumped to #1 on the App Store — the market rendering its own verdict. Anthropic CNBC CNBC Euronews Yahoo Finance

OpenAI’s agentic commerce ambitions have hit their own friction. After launching in-app shopping to much fanfare, the company pulled back its “Instant Checkout” feature — scaling down to referral links rather than native transactions. According to Forrester, direct buying within answer engines remains consumers’ least-adopted AI use case. link