AI Infrastructure Surge
We’re seeing a specific pattern across the largest tech platforms: AI is no longer just a feature being added to products. It is becoming measurable demand — in capital spending, cloud capacity, endpoint hardware, storage growth, token limits, and local model footprints.
Start with infrastructure. CNBC reports Alphabet raised its 2026 capital expenditure guidance to as much as $190 billion, with 2027 spending expected to increase significantly. Amazon is showing the same build cycle. Quartz reports AWS revenue rose 28% year over year to $37.6 billion, its fastest growth in fifteen quarters, while Amazon’s quarterly capital expenditures reached $44.2 billion, explicitly tied to AI infrastructure investment.
That tells us AI demand is not theoretical. The largest platforms are committing balance-sheet-level spending to support it.
But the signal is not only in hyperscale data centers. It is also moving closer to the client environment. TechCrunch reports Apple delivered $8.4 billion in Mac revenue in its March quarter, with executives pointing to demand from customers running AI workloads locally. And Blocks & Files reports Backblaze’s B2 cloud storage business grew 24% year over year to $22.4 million, with AI-related bookings called out as a meaningful driver.
So AI is pulling demand in two directions at once: upward into cloud infrastructure, and outward into endpoints and storage.
The constraints are becoming more visible too. ITPro reports Anthropic expanded Claude Code usage limits after a compute partnership adding more than 300 megawatts of capacity — over 220,000 NVIDIA GPUs — and raised API rate limits dramatically, including input token limits moving from 30,000 to 500,000 tokens per minute for some tiers. That is the meter becoming explicit.
And the most everyday signal may be the most operationally important. The Register reports Chrome can automatically download a 4GB local model file — Gemini Nano — onto user machines unless settings or enterprise policy disable it.
Put those pieces together and the signal is clear: AI is becoming a managed consumption layer. It shows up as cloud capex, endpoint demand, storage growth, token ceilings, and local files on machines your clients may not even realize are changing.
Control Layer Wins
The mechanism is not simply that organizations want AI. It is that they want measurable output from AI before they have the operating model to control the cost, quality, permissions, and handoffs behind that output.
PYMNTS reports that 34% of CFOs at billion-dollar companies cite increased output as the main reason for adopting AI, with productivity taking priority over “cool tools.” But only 12% say they feel very prepared for the workforce changes AI will bring.
That gap matters. Leadership is asking AI to produce measurable gains, but the organization often has not defined who approves usage, who measures consumption, what quality standard counts as acceptable output, when a human must intervene, or who owns the cost when automation expands beyond the original plan.
So vendors are turning operational control into the product.
VentureBeat reports Anthropic is rolling out Claude Managed Agents features such as “dreaming,” outcomes grading, and multi-agent orchestration. The important point is not the feature list. It is the packaging. Anthropic is trying to convert one-off AI use into repeatable work: tasks specified in advance, outputs graded against rubrics, and execution delegated across specialized agents.
That is what happens when AI moves from experimentation to expected business output. The value shifts from access to the model toward the system that can define, meter, evaluate, and constrain the work.
ServiceNow makes the same positioning explicit in its own newsroom release: enterprises have “AI chaos,” and the answer is a governed platform for autonomous work. In other words, the control layer becomes the sellable layer.
Put those pieces together and the mechanism is straightforward: as AI becomes expected infrastructure, the bottleneck moves from “can we use it” to “can we govern it.” The provider that supplies the control layer owns the meter. The provider that does not is left supporting usage, cost, and behavior it cannot see.
MSP Liability Shift
Once automation moves into the everyday tools your clients already run, the MSP problem changes. It is no longer just “can we support the application.” It becomes “can we govern what the automation is allowed to do, and can we prove we governed it.”
Microsoft is already making that shift visible. Thurrott reports that Microsoft Agent 365 is now generally available as an agent management layer, designed to let admins discover, monitor, and control both cloud-based and local AI agents. The important detail is the control model. Microsoft is tying agent oversight into Intune policy, including the ability to block or restrict classes of agents and sync in agents from other ecosystems.
That tells you where the industry is heading. Agents are becoming a managed estate. They are not just features inside applications. They are autonomous behaviors that have to be inventoried, permissioned, monitored, and constrained, the same way MSPs already manage identities, endpoints, configurations, and access policies.
Google shows the same shift at the endpoint level. Thurrott reports Google clarified that Chrome has been downloading on-device Gemini Nano models, consuming up to four gigabytes of local storage, to power AI features such as summarization, rewriting, tab organization, and certain security checks.
Again, the key question is not whether the feature is useful. The operational question is whether the MSP knows it is there, whether policy allows it, whether the client approved it, and whether the configuration can be verified.
That is the consequence for MSPs: AI issues will not arrive as clean “AI tickets.” They will arrive as storage growth, slow endpoints, unexpected usage, unauthorized workflow behavior, data-handling concerns, bad automated output, or a user asking why a tool behaved differently than expected.
The exposure chain is direct. If an AI feature consumes tokens, compute, storage, or endpoint resources without a defined approval path, the client experiences it as cost, performance impact, or uncontrolled data handling. If an agent acts outside its intended scope, the failure is not just “AI made a mistake.” It is a permissions failure, a workflow control failure, or an audit failure.
And if the MSP contract does not say who approves AI consumption, who sets caps, who owns policy exceptions, and who pays for overages, the MSP gets the escalation before anyone has a clean legal answer.
That is where the meter becomes the business model.
The MSP that can show last month’s AI consumption, identify which users and tools drove it, prove which agents were allowed to act, document which local models were deployed, and tie those controls back to contract language can price AI as governed infrastructure.
The MSP that cannot measure those things is left supporting invisible consumption inside agreements priced for seats, tickets, and devices.
Why Do We Care?
Platform vendors are establishing default control architectures inside managed client environments right now — not as a roadmap item, but as shipping behavior — and the billing meter has already shifted from seats to compute, tokens, and local model footprints. MSPs without a governance layer between that infrastructure and the client are holding liability proximity to a question that has no legal answer yet, on contracts priced for a model that no longer exists.
What to Consider
- Pull current contracts and identify any that lack AI consumption carve-outs. Token overages, storage growth driven by AI workloads, and compute spikes are not covered by flat-fee language written before 2024.
- Build a consumption monitoring capability — even a basic one — before positioning it as a service. The credibility gap between “we can govern AI spend” and “here’s your actual token consumption last month” is the difference between a retained client and a churned one.
- Don’t sell AI enablement. The hyperscalers have that covered at a price point no MSP can match. The defensible position is AI accountability — policy, audit, reporting, and liability proximity management. Frame it that way in client conversations now, before a competitor does.
If this trend continues, by 2027, the margin difference in managed AI services will not come primarily from resale discounts. It will come from whether the MSP can prove control over consumption and exposure. A provider that can report token use, storage growth, local model deployment, agent permissions, policy exceptions, and workflow output will have the basis for carve-outs, usage-based pricing, and governance fees. A provider that cannot will absorb AI-related tickets and disputes inside flat-fee contracts built for a pre-AI operating model.