The State of Cloud Development Environments in 2026
A synthesis of the published research on this category: where the reports contradict each other, which widely-quoted figures have no primary source, and what nobody has measured at all.
Compiled July 2026 from sources indexed on our sources page, each re-checked against the publisher. No new data collection.
What the Surveys Agree On
Two findings survive contact with every instrument, across separate populations and separate publishers.
Adoption is close to universal
The Stack Overflow Developer Survey 2025 reports that 84 percent of respondents use or plan to use AI tools in their development process, up from 76 percent in 2024, with 33,662 respondents answering that question. JetBrains State of Developer Ecosystem 2025 (October 2025, n = 24,534 across 194 countries) puts regular AI tool use at 85 percent, with 62 percent relying on at least one AI coding assistant, agent, or editor. Google Cloud's 2025 DORA report announcement (September 2025, nearly 5,000 respondents) reports 90 percent using AI at work.
Three publishers, three sampling frames, three instruments, all above 80 percent. That convergence is the strongest single claim available here, which is why the rest of this page is about what does not converge.
Trust is not
Every survey that asks about confidence finds a substantial minority that does not have it. Stack Overflow 2025 records 3.1 percent who highly trust the accuracy of AI output and 29.6 percent who somewhat trust it, against 45.7 percent who somewhat or highly distrust it, and names "AI solutions that are almost right, but not quite" as the top frustration for 66 percent of developers. Positive sentiment toward AI tools fell from above 70 percent in 2023 and 2024 to 60 percent in 2025. DORA finds 30 percent reporting little or no trust in AI-generated code.
Near-universal adoption paired with a durable trust deficit is what makes an isolation boundary a purchasing decision rather than a preference. Code that a large minority of its own users will not vouch for still has to run somewhere. GitHub Octoverse 2025 (October 2025, platform telemetry, vendor-published) gives the volume: 180 million-plus developers, 36 million joining during 2025, and public repositories importing LLM libraries up 178 percent year on year to more than 1.13 million.
Where the Surveys Disagree
Real disagreements, not rounding. Where two reports contradict, we present both and say we cannot resolve it. We do not average them.
Three adoption headlines, three different questions
84, 85 and 90 percent look like one finding. They are answers to three questions. Stack Overflow asks about tools respondents "use or plan to use" - a category that includes people who have not started. JetBrains asks about tools used "regularly". DORA asks about using AI "at work".
Note the direction that implies. The broadest question should return the highest number, and it returns the lowest. That points at a population difference rather than a wording artifact: DORA's respondents skew toward organizations engaged enough with delivery practice to answer a DORA survey, JetBrains' toward users of JetBrains products. Which figure best describes the working developer population is not resolvable without a shared sampling frame, and nobody has one.
The trust gap: 46 percent against 30 percent
In the same year, Stack Overflow measured 45.7 percent distrusting the accuracy of AI output while DORA measured 30 percent reporting little or no trust in AI-generated code. That is a 16-point spread on what reads like one construct. "Accuracy of the output" and "trust in the code it generates" are not identical questions and the response scales differ, which is very likely part of the gap. It is not the whole gap, and nobody has published a reconciliation. Both figures are correctly reported. Averaging them would produce a number that no instrument measured.
Belief against measurement
DORA reports that more than 80 percent of respondents believe AI has increased their productivity. The only randomized controlled trial in the area found the opposite. METR's July 2025 study put 16 experienced open source developers through 246 tasks on repositories they already knew well and measured them 19 percent slower with AI, while they predicted beforehand they would be 24 percent faster and believed afterwards they had been 20 percent faster.
That contradiction is now contested by one of its own parties. In a February 2026 update, METR states it believes developers are likely more sped up by AI in early 2026 than its early-2025 estimates suggested, and that it is abandoning task-level randomization. The reason given is selection bias: 30 to 50 percent of developers told the team they were declining to submit some tasks because they did not want to do them without AI, which removed the tasks where AI helped most. Belief and measurement still disagree, but the measurement side is weaker than it was in 2025, and it is weaker because its authors said so.
The denominator problem
Stack Overflow's 2025 survey has tens of thousands of respondents overall, but the AI usage question was answered by 33,662 of them, a 68.7 percent response rate published alongside the figure. Coverage routinely attaches the 84 percent to the full survey base instead. A smaller error than the ones below, and the same class of error: a number copied without the conditions that make it mean something.
Figures in Wide Circulation With No Traceable Primary Source
Each appears in published articles about this category as though established. Each traces to nothing, to a vendor's own assertion, or to a source that has since superseded it.
| The claim as it circulates | What it actually traces to | Why we will not repeat it |
|---|---|---|
| "Developers lose about 2.2 hours a week to local environment problems" | A 2023 blog post by an author who discloses membership of the Gitpod community team, attributing the number to Gitpod's chief executive. | No instrument, no sample size, no methodology, and a commercial interest in the answer. It is a vendor assertion quoted by a vendor-affiliated author, and it is the single most-repeated number in CDE marketing. |
| "AI makes developers 19 percent slower" | METR, July 2025. The figure is real and correctly reported. | Quoted uncaveated as a fact about the present, it is a superseded estimate. METR's February 2026 update says it now believes developers are more sped up than that number implies, and that it is abandoning the design that produced it. Cite the update with the study. |
| "The average data breach costs $4.88M" | IBM's 2024 Cost of a Data Breach report. The 2025 edition puts the global average at $4.44M and the United States average at $10.22M. | Two errors stack: the year is wrong, and it is an all-causes average across every breach type. Not a lost-laptop figure, not a development-environment figure, so it does not belong in a CDE business case. |
| "The 2026 Developer Survey found..." | The 2025 Stack Overflow results, relabelled. The 2026 survey opened in June 2026 and results were not published as of this review. | A year label is not decoration in a field this fast-moving. Anything presented as a 2026 developer survey finding right now is a 2025 finding wearing a new date. |
| "CDEs cut onboarding time by N percent" | In every case we followed, a vendor's own model or a single unnamed customer anecdote. We found no independent, published, before-and-after measurement of onboarding time across a CDE migration. | Onboarding is the most defensible qualitative argument for this category. Attaching a fabricated precision to it makes the argument weaker, not stronger. |
| "MicroVMs boot in under 200ms" | Firecracker's own SPECIFICATION.md commits to 125 ms or less from the InstanceStart API call to the start of the guest /sbin/init process, and 5 MiB or less of VMM memory overhead for a microVM with 1 vCPU and 128 MiB of RAM. | The specification figure is fine because it states its conditions. The rounded versions circulating without them get compared against workspace start times that include image pull, clone, and dependency install. |
What Nobody Has Measured
We went looking for each of these and did not find it. Saying so is more useful than filling the gap with an estimate.
Cost per successful agent task
Vendors publish cost per hour or per sandbox. Nobody publishes cost per accepted change. Failed attempts bill exactly like successful ones, so a per-hour price cannot tell you what the work costs.
CDE carbon impact, before and after
No published before-and-after measurement of moving developers from laptops to cloud environments exists. The nearest peer-reviewed work, Farthing, Langner and Trenbath (NREL, 2018, n = 4), found zero-client computing used 119 percent more power than laptops once data centre power was counted. See GreenOps.
Open-source project adoption
The onboarding argument should apply most strongly to drive-by contributors, and vendors offer free tiers to open-source projects. We found no published account of a major project moving its contributor workflow to a CDE with contributor numbers on either side of the change.
Self-hosting cost in engineer-days
A self-hosted control plane means a server, a database, an ingress, a template library, and an upgrade cadence. Nobody publishes what that costs in staffing. It is the largest unpriced line in every build-versus-buy comparison we have read.
Break-even seat count
The seat count at which self-hosting beats per-seat SaaS is the first question a platform team asks and the one the public record does not answer. Several vendors here publish no per-seat price at all, which makes it unanswerable from outside.
Whether isolation strength changes outcomes
That unreviewed generated code needs a hardware boundary is an argument from threat model, and we find it persuasive. It is not an argument from incident data: no published dataset compares breach rates across container, gVisor, and microVM agent workloads.
What Changed in the Category During 2026
Four names a 2024 shortlist would certainly have contained are, as of July 2026, in a materially different position. This is the practical reason to distrust an old comparison article.
| Change | Date | What the primary source says |
|---|---|---|
| Daytona left the category | Announced April 2025; source closed June 2026 | The daytonaio/daytona README states: "This repository is no longer maintained. As of June 2026, Daytona's core development has moved to a private codebase. This repository will receive no further updates, fixes, or releases." It is not archived. Its $24M Series A announcement (Daytona, 02-05-2026) is titled "Give Every Agent a Computer". An agent runtime company now, not a CDE company. |
| OpenAI agreed to acquire Ona (formerly Gitpod) | 06-11-2026 | Ona published "Ona is joining OpenAI" on its own blog. As of this review the transaction is announced, not confirmed closed. Anyone evaluating Ona is evaluating a product whose ownership is in transition and whose roadmap under that ownership is not yet published. |
| Microsoft Dev Box entered maintenance mode | Notice live as of July 2026 | Microsoft Learn carries this on the product overview: "Dev Box is now in maintenance mode, with no additional features planned. Microsoft's investments for developer cloud environments are focused on Windows 365." It adds that customers "should consider Windows 365 as the recommended path forward" and that Dev Box "will remain supported for existing usage". |
| Coder scheduled Tasks for removal | ESR from 06-02-2026; removal from v2.37 on 09-01-2026 | The Coder documentation states: "Starting June 2, 2026, Coder Tasks will move to a 12-month Extended Support Release (ESR) for Premium customers. Tasks will be removed from new Coder releases beginning with v2.37 (September 1, 2026) and will only be available via the ESR during the support period." Coder Agents is named as the long-term replacement. See our Coder page for what that means for a migration timeline. |
The pattern worth naming
Three of those four are movement toward agent infrastructure; the fourth is a vendor folding a developer-workstation product into a general virtual-desktop one. The category is not consolidating around a winner, it is splitting. Human workspaces drift toward virtual desktops and portable environment definitions, while the money and the engineering attention move to running unreviewed generated code safely. That split is what a two-year-old shortlist misses.
Method Note
What this is
A synthesis of already-published sources, each named, dated, and linked. Selection started from the sources indexed on our sources page. Every figure was re-checked against the publisher's own page in July 2026, and anything that did not survive that check was cut rather than softened.
What it is not
Not original research: no survey run, no benchmark executed, no interviews conducted. Not a market-share estimate, a vendor ranking, or a forecast. Not exhaustive - a report we could not reach is absent rather than judged.
When it gets re-checked
Quarterly, and immediately on two events: publication of the 2026 Stack Overflow Developer Survey results, and the closing or abandonment of the OpenAI acquisition of Ona. Found a stale figure or a rotted link? Tell us and we will correct or remove it.
Related Reading
The underlying evidence, and the decisions it feeds
Sources and Citations
Publisher, year, and link for every figure on this site
Which Isolation Boundary Do You Need?
The decision the trust deficit above turns into a purchase
Local vs Cloud Development
Where the evidence runs out and argument has to carry it
GreenOps
The sustainability figures we refuse to repeat, and the study that cuts the other way
CDE Tools Comparison
Every platform side by side, including those that changed category
Cost Analysis
Running the economics on your own utilization, since no break-even figure is published