Explosive AI Demand Drives Big Tech Cloud Growth

Big tech earnings underscore an undeniable AI boom, with cloud revenues surging due to insatiable demand for compute and tokens. Google Cloud grew 63% YoY, backed by a $460B order backlog (up from $240B) and 16B tokens processed per minute (up 60% QoQ). CEO Sundar Pichai noted, "Our enterprise AI solutions have become our primary growth driver for cloud for the first time in Q1," but admitted compute constraints limited revenue. Google raised CapEx guidance to $180-190B yet spent only $35.7B in Q1, signaling discipline amid GPU shortages.

Amazon's AWS hit 28% YoY growth ($152B ARR), accelerating from 2023 lows, fueled by OpenAI and Anthropic partnerships. CEO Andy Jassy highlighted custom Trainium chips: "As best as we can tell, our custom silicon business is now one of the top three data center chip businesses in the world." Q1 CapEx reached $43.2B toward a $200B annual target, nearly matching operating income and squeezing free cash flow to $1.2B. Microsoft Azure grew 39%, in line with forecasts, with Copilot at 20M paid seats (up from 15M). Satya Nadella downplayed lost OpenAI exclusivity: "We have a frontier model royalty-free with all the IP rights that we will have access to all the way to 32." Meta posted 33% revenue growth but faced stock declines after raising CapEx to $145B, with CFO Susan Li admitting, "We have underestimated our compute needs."

These results refute AI bubble skepticism—Sheharyar Khan noted Google's 63%, Azure's 40%, Meta's 33%, AWS's 28%—while highlighting uniform compute bottlenecks across hyperscalers.

Agent Evolution: From Weights to Harness Engineering

Agent progress has shifted beyond model scaling. Akshay outlined three phases: (1) Weights—bigger models via scaling laws, RLHF; (2) Context—prompt engineering, RAG, chain-of-thought for same-model variability; (3) Harness—persistent memory, reusable skills, sandboxes, protocols (MCP, A2A), observability. "The model is no longer the sole location of intelligence. It sits inside a harness," enabling reliability without model changes. Example: Coding agent with harness uses persistent repo context, skill files, failure handling—versus fragile prompts.

Sam Altman emphasized inseparability: In a Ben Thompson interview, he said, "Hard to overstate how critical the harness is. I no longer think of the harness and the model as these entirely separable things... I don't always know how much credit was it the model that's amazing or the harness that's amazing?" 2025's agent explosion combined Opus 4.5/GPT-5.2 with harnesses like Claude Code and OpenAI Codex. Open Claw democratized this but required builders to handle prompts, tools, loops, state, errors, deployment—akin to 1970s hobbyist kits like CompUKit UK101, per Anders Carlson's LinkedIn post on soldering bare boards.

Harness-as-a-Service: Scalable Agent Runtimes Emerge

A new category—"Harness-as-a-Service" (HaaS)—abstracts agent runtimes like AWS does compute. Cursor SDK offers local hackable agents or managed cloud ones, handling sandboxing, computer use, GitHub integration. Li Robinson: Build with any model, ship products. Recent launches: OpenAI agents SDK update, Anthropic Claude managed agents, Microsoft Foundry hosted agents—Nadella: "Every agent will need its own computer... dedicated enterprise-grade sandbox with durable state, built-in identity and governance."

HaaS provides sandboxed execution, state persistence, monitoring—turning LLMs into workers. Benchmarks show performance gains; apps proliferate in coding, IT triage, workflows. This layers atop prior phases, moving center of gravity outward. Agent OS (tool-agnostic system) complements by enabling adaptable OSes post-HaaS.

Implications: From DIY to Production Agents

HaaS ends hobbyist era, enabling non-experts to deploy reliable agents without wiring loops or managing infra. Trade-offs: Vendor lock-in vs. customization; costs scale with usage like cloud. Yet, it accelerates agentic apps—Cursor demos reveal rapid prototyping to production. Big tech's compute surge funds this infra, positioning HaaS providers as picks-and-shovels winners amid token droughts.

Key Takeaways

  • Track hyperscaler CapEx and backlogs (e.g., Google's $460B) as leading AI demand indicators—demand outpaces supply.
  • Prioritize harness over models: Build persistent memory, sandboxes, protocols for 10x reliability on same LLMs.
  • Adopt HaaS early: Test Cursor SDK for coding agents, Anthropic/Microsoft for enterprise—handles 80% boilerplate.
  • Layer phases: Weights + context + harness = production agents; ignore any for fragility.
  • Monitor agent runtimes like cloud: Sandboxing, state, governance prevent hallucinations/escalations.
  • Explore open tools like Agent OS post-HaaS for custom OSes.
  • Bet on infra plays: Custom silicon (Amazon Trainium), partnerships (OpenAI on Bedrock) yield moats.
  • Demand proof in earnings: 20M Copilot seats = traction; scale to Office 365 levels needed.
  • Avoid DIY pitfalls: Open Claw great for prototypes, but HaaS for shipping.
  • Compute constraints universal—optimize tokens/min (Google's 16B) via efficient harnesses.