When David Crawshaw sat down to write his April 22 blog post, he wasn't aiming for subtlety. The hyperscalers, he argued, have built abstractions that are fundamentally "the wrong shape." Virtual machines sized in rigid CPU-to-RAM ratios. Remote block devices imposing IOPS overhead where local SSDs should fly. Egress fees taxing every byte that crosses the perimeter. And Kubernetes—the orchestration layer everyone reached for to smooth over these friction points—can't paper over what Crawshaw sees as foundational missteps from a different computing era.
That same day, his company, exe.dev, announced a $35 million Series A led by Amplify, CRV, and HeavyBit. The pitch: a cloud built on its own hardware, racking machines in data centers and offering developers CPU and RAM pools they can slice into as many virtual machines as they need. No pre-sizing. Local NVMe with asynchronous replication. Built-in TLS and anycast ingress.
The round places exe.dev in a widening field of alternative cloud providers pushing back against AWS, Microsoft Azure, and Google Cloud—not on scale (they can't win that fight) but on developer experience, cost legibility, and architectural philosophy. Whether this cohort represents a genuine shift or just well-funded noise depends on forces now coming to a head: regulatory pressure dismantling egress fees and switching barriers, AI workloads outgrowing primitives the big three weren't designed for, and perhaps most intriguingly, a developer mood that's soured on complexity and billing surprises.
The Hyperscale Fortress, and Its Cracks
Cloud infrastructure services pulled in roughly $119.1 billion during the fourth quarter of 2025, according to Synergy Research estimates published in February 2026. Full-year 2025 revenue approached $419 billion. AWS, Azure, and Google Cloud captured somewhere between 62 and 63 percent of that across multiple quarters, with AWS holding a lead that's eroded modestly over time—down from higher shares in prior years to somewhere in the 29 to 33 percent range, depending on which quarter and analyst series you consult. Azure and Google Cloud have been clawing back share, growing faster, but the trio's grip remains firm.
Gartner projected in November 2024 that end-user spending across all public cloud service layers would hit $723.4 billion in 2025, up from $595.7 billion the year before. Canalys, analyzing Q4 2024 data in early 2025, forecast 19 percent global growth in cloud infrastructure services for the year. Other analysts, cited in March coverage, suggest 2026 could surpass $500 billion, with year-over-year growth near 27 percent—much of it driven by generative AI compute demand.
Yet beneath those numbers, something is shifting. In January 2025, the UK's Competition and Markets Authority issued provisional findings calling out egress fees and licensing practices as obstacles to multicloud adoption and switching. By April 2026, the CMA announced specific actions targeting AWS and Microsoft on egress and interoperability for UK customers. Across the Atlantic, the EU's Data Act mandated cloud switching obligations that went live in September 2025, with a complete ban on switching and egress charges scheduled for January 12, 2027. Even the U.S. Federal Trade Commission, in a November 2023 request for information on cloud computing, flagged egress fees and licensing lock-in as areas of concern. Regulatory scrutiny, it turns out, is not a European peculiarity.
Three Forces Converging
The openings for new entrants come from at least three directions. First: cost transparency. Hyperscaler pricing is Byzantine—line items for compute, storage, networking, and data transfer that compound in ways developers often discover only when the invoice arrives. A 2019 AWS architecture blog post detailed the labyrinth of data transfer costs across services and paths, illustrating how even well-architected workloads can rack up unexpected charges. Egress fees are the most visible pain point. AWS historically charges $0.09 per gigabyte for the first 10 terabytes monthly of internet egress, softened by a 100 GB monthly free tier introduced in November 2021 and an additional 1 TB via CloudFront.
Cloudflare positioned its R2 object storage as a zero-egress alternative, with pricing pages updated as recently as late 2025. The Bandwidth Alliance, launched in 2019 and still operational, offers zero-cost egress between partners like Backblaze B2 and Cloudflare. Providers such as OVHcloud and Hetzner bundle generous bandwidth into their tiers—though Hetzner announced price increases of up to 37 percent effective April 1, 2026, citing global hardware cost surges. A reminder, perhaps, that "cheap" is never entirely divorced from input economics.
Second: architectural misalignment. Crawshaw's critique centers on the gap between hyperscaler primitives and what modern applications actually need. Remote block storage—even AWS's io2 Block Express volumes offering up to 256,000 IOPS and 4,000 MB/s throughput with sub-millisecond latency—still introduces network coordination overhead that local NVMe SSDs sidestep entirely. exe.dev's design reflects a bet that developers want lower-level control without operational headaches. CPU and RAM pools let you spin up dozens of lightweight VMs without pre-sizing each one. Built-in TLS, an IAM proxy, and anycast ingress come standard.
Fly.io made a parallel bet in January 2026 with Sprites—persistent virtual machines designed to keep AI agents "warm" with state intact. The CNCF's 2026 annual survey, published in January 2026 and covering 2025 adoption patterns, found that 98 percent of surveyed organizations report using cloud-native technologies in some capacity, with 82 percent of container users running Kubernetes in production. Ubiquitous, yes. But the same survey hints at a tension: Kubernetes is also complex. New compute primitives—whether Sprites, exe.dev's pooled VMs, or GPU-specific offerings—aim to abstract that complexity for workloads that don't need full orchestration.
Third: power and capacity constraints. The U.S. Energy Information Administration forecast in January 2026 that the next four years would see the strongest electricity demand growth since 2000, with data centers as a primary driver. CBRE reported in March that Northern Virginia's total data center inventory exceeded 4.0 gigawatts by the end of 2025, with more than 1 GW delivered during the year alone. Power delivery is batching. Timelines are extending. Axios, citing Sightline data in February, noted a surge in local moratorium proposals on new data center construction—even as 2026 is on track to surpass 2025's record additions.
For companies building custom infrastructure, those constraints are double-edged. Securing power and rack space is harder, but if you manage it, you may sidestep some of the hyperscaler capacity crunches that push prices or extend availability windows.
The Challengers Taking Shape
exe.dev isn't operating in a vacuum. Oxide Computer Company raised a $100 million Series B in July 2025 to scale its "cloud computer"—a rack-scale system integrating hardware and software to deliver cloud-like primitives on-premises or in colocation facilities. The pitch: hyperscaler developer experience without the hyperscaler, control and cost predictability for enterprises that want both. Red Hat OpenShift demos across 2024 and 2025 showcased Oxide racks in enterprise settings. Oxide has also positioned itself as a sovereign cloud option for organizations navigating data residency and geopolitical pressures. Gartner estimated in February 2026 that European sovereign IaaS spending would climb from $6.9 billion in 2025 to $12.6 billion in 2026—an 83 percent jump—with global sovereign cloud IaaS hitting $80.4 billion in 2026. A separate Gartner note from April suggests up to 20 percent of IaaS workloads may shift from global to local providers by 2027, driven by geopolitics and data locality demands.
The GPU cloud segment tells a similar story. CoreWeave, positioning itself as a hyperscale alternative for AI training and inference, disclosed in March 2026 investor materials a revenue backlog of $66.8 billion as of December 31, 2025, with plans to add 5 GW of capacity by 2030. NVIDIA has made strategic investments and granted early access to next-generation chips, cementing CoreWeave's role in the GPU supply chain. Crusoe raised an initial $1.375 billion in a Series E at more than $10 billion valuation in October 2025, betting on modular data centers and an "AI factory" model that pairs compute with on-site power generation, often at stranded energy sites. Its Texas campus went live last year. Voltage Park, which announced in February 2024 that it had deployed 24,000 H100 GPUs, lists pricing as low as $1.99 per hour for an H100 SXM5 on its current site—a headline rate undercutting many hyperscaler spot market floors.
Historical precedents exist for these moves, under the right conditions. Dropbox disclosed in March 2016 that it had built Magic Pocket, a custom storage backend, to shift the majority of user file storage off AWS. The engineering effort was substantial—the company leveraged shingled magnetic recording and custom hardware to drive down costs. By 2019, Dropbox was still using AWS for regional deployments and certain services, but the core storage economics had shifted in its favor. 37signals, the company behind Basecamp and HEY, published a "Big Cloud Exit FAQ" in December 2023 claiming it had spent $3.2 million on cloud bills in 2022 and projected multi-million-dollar annual savings by owning hardware. Founder DHH emphasized that cost was only part of the calculus—control, reliability, and an ideological preference for owning infrastructure mattered, too.
Caveats abound. Dropbox and 37signals had mature, stable workloads and the engineering capacity to manage infrastructure. For most startups, the hyperscalers' marginal cost and instant scale remain hard to beat. But the existence of these examples—and the willingness of investors to back hardware-owning challengers—suggests a growing belief that the hyperscale model isn't universally optimal. Or perhaps more modestly, that there's room at the margins.
What Comes Next
Regulatory momentum will likely compress egress fees and switching friction, at least in the EU and UK. The Data Act's January 2027 deadline is firm. UK CMA actions announced in April 2026 indicate that AWS and Microsoft are already making concessions for UK customers. Whether those changes extend globally is an open question—U.S. regulators have signaled interest but haven't advanced formal rulemaking. Still, the precedent is set.
Power and capacity constraints will shape where and how new providers compete. The EIA's 2026 outlook underscores that data center electricity demand is accelerating, not plateauing. Siting challenges, local moratoria, grid investment timelines—these mean that securing power is both a competitive moat and a potential bottleneck. Providers like Crusoe, betting on modular facilities at stranded energy sites, are attempting to sidestep those constraints. Hyperscalers are investing heavily in their own generation and transmission. But they face the same physical limits.
Developer experience will likely bifurcate. Kubernetes remains the backbone for cloud-native workloads, as the CNCF's survey confirms, but abstraction layers tailored to specific use cases—AI agents, real-time data processing, edge compute—are proliferating. exe.dev's CPU and RAM pools, Fly.io's Sprites, GPU clouds' per-second billing: all reflect a belief that one-size-fits-all compute is giving way to more opinionated platforms. The hyperscalers will respond—AWS, Azure, and GCP have resources and reach no challenger can match—but the window for differentiation on developer ergonomics is, for now, open.
Make no mistake: AWS, Azure, and Google Cloud will remain dominant. The $419 billion in 2025 cloud infrastructure revenue and forecasted growth through 2026 are driven overwhelmingly by enterprises with sprawling footprints, AI training runs needing thousands of GPUs for weeks, and regulatory or compliance demands that only the hyperscalers can satisfy at global scale. Dominance, though, isn't the same as universality.
The alternative cloud wave—whether developer-first platforms like exe.dev, sovereign on-prem racks like Oxide, or GPU neoclouds like CoreWeave—is carving out niches where cost, control, or architectural fit matter more than breadth. Crawshaw's $35 million Series A is a bet that enough developers are frustrated enough with hyperscaler abstractions to try something built from different first principles. Whether that bet pays off hinges on execution, timing, and the pace at which regulatory and market forces reshape the competitive landscape.
But the capital is flowing. The regulatory pressure is real. And the demand for alternatives is no longer just theoretical.
