If 2024 was the year cloud lock-in started losing — egress fees falling, the Data Act passing, the first open-source forks finding their footing — then 2025 was the year the consequences of an AI-driven infrastructure buildout collided with a maturing regulatory order. The two stories ran in parallel all year and increasingly intersected. The economics of AI inference reshaped what datacenters get built and where; the regulatory environment reshaped where data is allowed to live and who can be held accountable when it moves. Here is how the year looked from the infrastructure layer.

The Inference Economy Arrived

The year opened with a jolt. In late January, the release and rapid attention around DeepSeek’s reasoning model reframed a question that had been simmering since the start of the generative-AI buildout: how much does inference actually have to cost? The market reaction was sharp, and the infrastructure-economics conversation it triggered outlasted the initial shock. The lasting point was not about any single model. It was that the cost structure of serving AI — not training it — would determine the shape of the infrastructure layer, because inference is the part that runs continuously, at scale, for as long as a product exists.

That reframing rippled through the rest of the year. It put pressure on GPU pricing assumptions, sharpened interest in inference-optimized hardware and serving stacks, and made the placement of inference workloads — centralized in large GPU clusters versus pushed toward the edge — a live architectural debate rather than a theoretical one.

Silicon and the Physical Limits

The defining constraint of 2025 turned out not to be chips but power. As the buildout accelerated, the bottleneck shifted to datacenter power availability: grid interconnection queues stretched out for years, utilities struggled to commit capacity, and operators turned to power purchase agreements, on-site generation, and revived nuclear arrangements to secure the megawatts that AI-class clusters demand. The story of cloud capacity in 2025 was increasingly a story about energy infrastructure, and that is a domain that moves on the timescale of utilities and regulators, not software releases.

Rack densities climbed with the new accelerator generation, and the cooling required to handle them followed. Over the course of the year, liquid cooling moved from specialist deployments into mainstream practice for high-density AI racks — direct-to-chip and immersion approaches that air cooling simply cannot match at the power densities involved. The physical plant of the datacenter, long an afterthought to most software teams, became a first-order design constraint.

Cryptography Quietly Advanced

While the AI buildout absorbed most of the attention, the post-quantum migration that began in earnest in 2024 continued steadily. In March, NIST selected HQC as a backup key-encapsulation mechanism, adding algorithmic diversity to the lattice-based standards finalized the previous year. This is the kind of progress that produces no market shock and no conference keynote frenzy, but it sets the timeline that security and infrastructure teams have to plan against. The migration to post-quantum cryptography is a multi-year program, and 2025’s standards work is what those programs will be built on.

Open Source Found Its Footing — and Reversed Course

The license-change wave that defined 2023 and 2024 produced one of the year’s more interesting reversals. In May, Redis returned to an open-source license, adopting AGPLv3 alongside its source-available options — an acknowledgment, however framed, that the fork ecosystem that grew up after the original relicensing had changed the calculus. The episode validated a pattern that became clearer through the year: when a widely depended-upon open-source project relicenses, a credibly governed fork under neutral stewardship is now a realistic outcome, and the existence of that option changes what the original steward can sustainably do.

On the regulatory front, the EU Data Act’s cloud-switching provisions reached their applicability milestone, and the long buildup to that deadline shaped procurement and architecture decisions across European enterprises throughout the year. The shift from anticipation to applicability was one of the quieter but more consequential developments of 2025, turning portability from an aspiration into an enforceable expectation.

The Year the Cloud Broke (Twice in Ten Days)

October provided the year’s most visible reminder that scale is not the same as invulnerability. Two major hyperscaler outages within a single week — one rooted in a regional control-plane dependency, one in a global edge layer — disrupted a wide swath of internet services and prompted a fresh round of failure-domain analysis. The lesson was familiar but worth restating: the failure domain of a cloud service is usually larger than the region boundary suggests, and redundancy aimed at the wrong failure domain is no redundancy at all.

Threads That Will Carry Forward

A few observations about where the year leaves the infrastructure layer, offered as observations rather than predictions:

  • Inference placement is unresolved. The question of how much AI serving lives in centralized GPU clusters versus distributed toward the edge is still open. The economics — latency, model size, power, data gravity — point in different directions for different workloads, and the industry has not converged on an answer.
  • Power is the binding constraint. Compute capacity will be gated by energy availability and grid access for the foreseeable future. Software efficiency and workload scheduling matter more in a power-constrained world than they did in a capacity-abundant one.
  • Sovereignty is now a design input, not a checkbox. Between the Data Act, national cloud initiatives, and AI-data residency concerns, where data is processed has become an architectural constraint that infrastructure teams design around from the start.
  • The open-infrastructure stack keeps maturing. The combination of forks, license reversals, and the VMware migration pressure of the past two years has strengthened the open-source private-cloud ecosystem rather than weakened it.

The dangerous thing about year-in-review pieces is the temptation to extrapolate confidently into the next year. The honest summary is narrower: 2025 made AI inference an infrastructure problem rather than a model problem, made energy the limiting resource, and continued the slow regulatory and open-source pressure on lock-in that the prior two years had set in motion. How those forces resolve is a story for the year ahead.

Further Reading

Back to Blog