The Week Google Rewired Everything

The headline is the hardware, not the model. Google didn't just release Gemini Spark this week — it announced that AI agents now run on always-on, persistent cloud VMs. That's the real story. For the first time, a frontier AI agent has a continuous home: it doesn't spin up when you call it, then disappear. It lives, waits, and acts — autonomously, around the clock. This is a meaningful architectural shift, not a product update.

The pricing tells you who Google is targeting. A $100/month Ultra tier for consumers and a $200/month enterprise plan are competitive when stacked against what organisations currently pay for automation tooling and human-hours on repetitive cognitive tasks. Google is pricing Gemini as infrastructure — the same category as compute, storage, and networking — not as a chatbot subscription. That framing matters enormously for how enterprise technology leaders should evaluate it.

Persistent agents change the security perimeter. When agents run continuously on cloud VMs — browsing, writing, calling APIs, sending emails on your behalf — your IAM policies become the last line of defence. An agent that lives 24/7 in your cloud environment has the same access footprint as an always-on employee with no off switch. Most enterprise IAM frameworks were not designed with this actor type in mind.

The governance issue is not theoretical. Google's move this week makes agent identity and access management a board-level infrastructure question, not a future concern. Enterprises that are still treating agentic AI as a pilot-phase experiment need to reset their timeline. The always-on agent is here, it's priced for enterprise scale, and it's running whether your governance team has caught up or not.

AI Mode is now the default. Google completed its rollout of AI Mode as the primary search experience this week, making the AI-generated summary the first and often only result a user sees. The implications are stark: 60% of searches now end without a single click to an external website. The search box no longer indexes the web — it answers questions about it. That distinction will reshape how information flows across the internet for the next decade.

The economics of web publishing are being restructured in real time. Publishers who built their traffic models on Google search referrals are watching that channel collapse. Enterprise content teams, investor relations pages, product documentation sites — anything built on the assumption that search drives discovery — needs to reckon with a world where the AI answers directly and the source link is optional. This is not a gradual trend. The transition from 10% to 60% zero-click happened in under 18 months.

Enterprise intranets are the next frontier. The same technology reshaping public search is being embedded into enterprise knowledge management. Microsoft Copilot, Google Workspace AI, and Glean are all racing to become the AI search layer inside organisations. The race is for the enterprise equivalent of the search box: the interface through which knowledge workers find, synthesise, and act on internal information. Whoever wins that position has enormous leverage over enterprise workflow.

For CIOs, this raises two parallel questions. First, how does your organisation's external web presence need to evolve when AI answers replace search results — and what does that mean for SEO strategy, knowledge base investments, and content governance? Second, which vendor do you trust to be your internal AI search layer, and what are the data residency and access control implications of that choice?

This is the most strategically significant acquisition in the AI toolchain to date. Stainless is not a consumer product — it's the SDK generation platform that OpenAI, Google, and dozens of other AI providers have relied on to build the developer libraries their customers use to integrate AI into applications. Anthropic acquiring Stainless for a reported $300M+ is the equivalent of a car manufacturer buying the company that makes the factory robots used by all its competitors. The toolchain war is now explicit.

What this means for vendor lock-in has moved upstream. Until now, AI vendor lock-in operated primarily at the model layer: switching from GPT-4 to Claude required re-prompting, re-tuning, and re-testing. Stainless moves the leverage point to the SDK layer — the code that wraps the API and sits inside enterprise applications. If Anthropic controls the SDK generator, it controls how easily developers can build for — or migrate away from — any AI provider. That's a different kind of competitive moat.

The developer community response has been pointed. Engineers who relied on Stainless-generated SDKs for multi-model integrations are now asking whether those tools will remain neutral or become subtly optimised for Claude. Anthropic has stated the platform will remain open, but the structural incentive runs the other way. Enterprise procurement teams should be having conversations with their development organisations now about SDK dependencies and migration risk.

For enterprise AI strategy, the key lesson is that the AI stack is consolidating faster than expected. The era of mixing and matching best-in-class components from competing providers is getting shorter. The major labs are racing to own not just the model but the development environment, the deployment infrastructure, and now the integration tooling. Portability clauses and multi-vendor architecture decisions made today will look very different in 18 months.

The number is staggering, but the breakdown is sobering. Gartner's latest AI spending forecast puts global AI-related expenditure at $2.59 trillion by 2027, representing 47% year-over-year growth. But when you disaggregate the figure, the enterprise share is smaller than the headline implies. The majority of current spending is hyperscaler infrastructure — data centres, chips, and cloud capacity — not enterprise AI deployment. The enterprise wave is coming, but it hasn't arrived at the scale the top-line number suggests.

Gartner places most large enterprises in the Trough of Disillusionment. After the peak enthusiasm of 2023–2024, organisations are confronting the gap between AI's demonstrated potential and their ability to capture it at scale. The barriers are not technical — they're organisational: data quality issues, workflow redesign debt, talent gaps, and governance frameworks that weren't built for AI-speed decision cycles. The technology is ready. The enterprise operating model often isn't.

2026 is the inflection year, but only for organisations that move now. Gartner's analysis identifies a cohort of enterprises — roughly 15% of the market — that are progressing through the trough into the Slope of Enlightenment: measurable productivity gains, repeatable deployment patterns, and AI-adjusted operating models. The differentiator in every case is the same: they started redesigning workflows before deploying tools, not after. Sequence matters more than technology selection.

The forecast also carries a warning about AI infrastructure costs. Enterprise GPU and inference costs are not declining as fast as adoption projections assume. Organisations that modelled AI economics on 2023 pricing are discovering that scaled deployment costs more than the pilot suggested. Rigorous AI unit economics — cost per inference, cost per workflow automation, cost per decision supported — need to be part of every enterprise AI business case.

When two Fortune 500 companies switch coding agents in the same week, it's a market signal. Both Microsoft and Uber announced this week that they were moving away from Claude Code as their primary AI coding assistant, citing performance differences on specific enterprise codebases and integration friction with internal developer tooling. The moves landed differently in each case — Microsoft towards its own Copilot stack, Uber towards a custom fine-tuned model — but the headline is the same: the coding agent market is live, competitive, and not yet locked in.

This is healthy market behaviour, but it carries a procurement lesson. The enterprises that will navigate the coding agent wars well are those that made tool selection decisions with explicit portability assumptions built in. Teams that integrated Claude Code at the infrastructure level — baking it into CI/CD pipelines, code review workflows, and developer onboarding — face higher switching costs than teams that kept the integration at the tool layer. Architecture decisions made during pilots have a way of becoming structural constraints at scale.

The performance gap between coding agents is real and task-dependent. No single agent is best across all enterprise codebase types, languages, and use cases. Microsoft's internal benchmarks reportedly favour Copilot on TypeScript and C# — unsurprisingly, given training data proximity. Uber's evaluation favoured a fine-tuned model on their specific microservices architecture. The implication for enterprise evaluation frameworks is that benchmark performance on public datasets is a poor predictor of performance on proprietary codebases. You have to test on your own code.

For enterprise technology leaders, the coding agent category deserves the same procurement rigour as any developer platform decision. That means: evaluation on internal codebases, clear performance metrics tied to developer productivity outcomes, contractual portability provisions, and a review cadence that matches the pace at which the category is evolving — which is currently quarterly, not annually.