NextFin News - Erik Bernhardsson, co-founder and chief executive of Modal Labs, framed the current moment as a "new era of software," and the clearest evidence he can point to is his own term sheet. Modal Labs is in talks to raise fresh capital at a valuation of about $15 billion, roughly three times the $4.65 billion price it secured in a $355 million Series C just four months earlier, as annualized revenue grew fivefold from about $60 million in September 2025 to roughly $300 million by May 2026. The question investors now face is whether that repricing reflects a durable shift in how software is built, or a capital-expenditure wave that will mean-revert once the AI coding boom slows.
The Claim and the Numbers Behind It
Bernhardsson's remarks, made in an interview on October 14, position artificial intelligence not as a cyclical upswing but as a regime change in software development. The timing is what makes the claim hard to dismiss. In May, the chief executive announced the Series C himself, writing that Modal had "raised $355 million after growing fivefold since September, surpassing $300 million in annualized revenue," with a "$4.65B post-money" valuation in a round led by General Catalyst and Redpoint Ventures, with Menlo Ventures and Accel joining as new investors. By late September, the company was already negotiating its next round at a price that would more than triple that figure.
A valuation moving from $4.65 billion to about $15 billion in roughly four months implies investors are underwriting more than the revenue Modal has already won. They are pricing a world in which AI agents become the default way software gets written, tested, and shipped. Modal's own usage metrics feed that thesis: the company says more than one billion sandboxes have been launched on its platform, which serves over 10,000 teams. Its customer list reads like a map of the AI economy — Anthropic, Meta, DoorDash, Cognition, the maker of the Devin coding agent, Suno, Physical Intelligence, Chai Discovery, Ramp, Lovable, and Quora.
The stakes extend well beyond one startup. If Bernhardsson is right that software development itself has entered a new era, the value created will accrue less to the model laboratories and more to the layer that runs the code those models generate. That is the bet embedded in Modal's valuation, and it is a bet the broader market is now being asked to price.
There is also a branding dimension to the moment that Bernhardsson has flagged separately. He has argued that the "era of boring brands is over" in enterprise infrastructure, pointing to a new generation of developer tools — Vercel and PlanetScale among them — that market themselves with the energy of consumer products. In a category where buyers are increasingly young engineers rather than procurement committees, a distinctive brand is no longer decoration; it is a distribution channel. That is a secondary thesis, but it helps explain why infrastructure companies are suddenly behaving like media companies.
Why the Bottleneck Moved From Training to Running
The mechanism behind the "new era" claim is concrete, and it is physical rather than rhetorical. AI coding agents do not merely suggest text; they execute code. They spin up environments, install dependencies, run tests, and iterate on untrusted output — thousands of times a day. Traditional cloud primitives were never built for that workload. A virtual machine that takes minutes to boot, or a container that must pull a multi-gigabyte image before it can run a single function, breaks the feedback loop that makes an agent useful.
Modal's answer was to rebuild the cloud from the bottom up. The company built its own file system, container runtime, scheduler, and container image builder — infrastructure most customers never see, but which delivers sub-second cold starts and the ability to scale to tens of thousands of concurrent sessions. For agent workloads, the critical feature is isolation: Modal runs untrusted generated code in gVisor-based sandboxes that block malicious system calls while still allowing GPU access when a task needs it. The platform advertises support for more than 50,000 concurrent sessions, a scale that matters when a coding agent serves thousands of developers simultaneously.
This is where the second-order effect kicks in, and it is the part most of the market is not pricing. The first-order story is familiar: AI makes developers more productive, so they use more compute. The second-order story is that productivity changes what compute looks like. When code generation becomes cheap, the scarce resource shifts from writing code to running, testing, and validating it. The winner is not necessarily the creator of the best model, but the layer that can execute untrusted code securely, instantly, and at a cost low enough that running it a thousand times is cheaper than a developer doing it once.
Capital is already voting on that shift. Baseten, another inference-infrastructure specialist, closed a $300 million Series E in January 2026 at a $5 billion valuation, with participation from Nvidia, and within months was reported to be raising again at a price roughly double that. The acceleration says more about the category than about any single company: money that spent 2024 and 2025 funding model laboratories is now chasing the layer that turns trained models into running products.
The demand signal behind the infrastructure rush is measurable. The market for enterprise AI coding agents has grown into an estimated $9.8 billion to $11.0 billion in annualized revenue as of April 2026. A survey of engineering leaders found that 90 percent report productivity improvements after adopting AI coding agents, with a net average gain of 19.3 percent. GitHub's Copilot, the category's largest distribution channel, has been used by more than 20 million people in total. And AI-authored code now accounts for roughly 22 percent of merged code across adopting teams.
"I think AI represents a new tool and it's actually really good, but I think to me, it's not necessarily fundamentally different than what's already been happening over the last 30 years with engineers being more and more productive," Bernhardsson said in a separate conversation about developer infrastructure. "If you look at the last 30 years, employment of engineers just keeps going up. There's more and more engineers in society."
That historical frame is the core of Bernhardsson's structural argument: AI does not replace developers; it multiplies them. He has argued that AI could lead to ten times as many developers, because the constraint on software creation has never been the number of people who can type code, but the cost of turning an idea into working software. In a related vein, he has said that "there's always more money in creating new markets" than in making existing workers more productive, and that the real gains will come from "a bunch of new jobs no one's even thought about."
The Economics of the New Layer
The bull case for infrastructure multiples rests on a simple arithmetic: if every line of AI-generated code must be executed, tested, and validated in a secure sandbox, then infrastructure consumption grows faster than the developer headcount that generates it. A tenfold increase in developers, or even a tenfold increase in code iterations per developer, compounds into a much larger increase in sandbox-hours. That is the operating-leverage story, and it is why revenue can grow fivefold in eight months rather than fivefold in five years.
But the arithmetic cuts both ways. Inference and code-execution costs are ultimately denominated in silicon and electricity, and both have trended down per unit of capability for a decade. If model efficiency improves faster than usage grows — if a task that costs a dollar to run today costs ten cents next year — then revenue per unit of economic value delivered compresses even as volume expands. The infrastructure layer can become a victim of its own success: the better it gets at making execution cheap, the harder it is to charge a premium for it.
This is the classic utility trap, and it is the reason the moat cannot be the hardware. Any specialist can rent GPUs; the defensible asset is the software layer above them — the scheduler that packs workloads efficiently, the image format that starts in milliseconds, the isolation model that lets untrusted code run safely, and the developer experience that makes all of it invisible. Bernhardsson has called the quality of the tool's ergonomics the company's "core competitive advantage," the reason customers pick Modal over cheaper alternatives. If that holds, pricing power survives commoditization. If it does not, the category converges toward cloud pricing, where margins are real but multiples are not.
Cyclical Wave or Structural Shift? The Counter-Thesis That Could Break the Trade
The modal valuation call — a $15 billion price on roughly $300 million of annualized revenue, or about 50 times revenue — only makes sense if the shift is structural. A cyclical read would say this is a capex wave: companies are overbuilding agent infrastructure today because funding is cheap and the narrative is hot, and spending will revert when the novelty fades or when inference margins compress.
The strongest evidence for the structural view is a change in the software primitive itself. Code execution as a service is not a feature that can be bolted onto a legacy cloud; it requires a different scheduler, a different isolation model, and a different image format. Once an AI company builds its product on a sandbox runtime, switching costs are high. There is a historical analog worth holding loosely: the platform-as-a-service wave of the early 2010s, epitomized by Heroku, which convinced a generation of developers that infrastructure should be invisible. That wave was real and durable in its impact — it changed developer expectations permanently — but the financial value did not accrue to the specialists. The hyperscalers absorbed the demand, and many PaaS brands were acquired at modest multiples. The lesson is not that Modal is Heroku; it is that changing developer behavior and capturing enterprise value are two different achievements, and history has rewarded the former less generously than its participants expected.
But the counter-thesis is serious and deserves its due weight. A working paper by Harvard economists Fiona Chen and James Stratton, analyzing roughly 300 million work events across more than 700 software firms from 2021 through March 2026, found that after AI agent adoption each worker added about 1,500 extra lines of code per month — code volume up roughly 30 percent, with 20 percent more commits and 23 percent more pull requests. Yet the study found no statistically significant increase in finished issues or completed epics. More code, in other words, did not translate into more shipped software. Review time per pull request rose 49 percent, to about 3.45 days, and revision rates nearly doubled. If that pattern holds, the productivity boom is an illusion of output: developers are generating more code to review, more code to debug, and more code to maintain, without moving faster from idea to shipped product. In that world, agent usage plateaus, inference demand growth slows, and the infrastructure multiples compress back toward historical software norms.
There is a second leg to the bear case: vertical integration. The largest customers of inference infrastructure are also the companies with the most to gain from owning it. Anthropic, Meta, and the other frontier laboratories already run their own training clusters; if their agent workloads grow large enough, building a proprietary execution layer becomes cheaper than paying a 50-times-revenue markup to a specialist. The same dynamic played out in the first cloud era, when the biggest workloads eventually migrated in-house or to the hyperscalers that could bundle them with storage, networking, and identity.
The falsifying signal is specific. Watch the share of merged code that is AI-authored, which currently sits around 22 percent. If that share stalls below its current level for two consecutive quarters while agent-seat adoption continues to rise, it would confirm the "output without shipped software" critique and undercut the structural thesis. A second signal: if the largest AI laboratories publicly announce proprietary agent-run-time platforms that replace third-party sandboxing, the vertical-integration bear case has begun to print.
Who Benefits, Who Is Exposed, and What to Watch Next
The implications diverge across time horizons. In the short term, the trade is liquidity- and narrative-driven: any company associated with AI inference, developer tooling, or secure code execution can reprice on the next funding headline, regardless of revenue. That momentum can persist well past the point where fundamentals justify it, and it can reverse just as fast.
Over the medium term, the question is revenue catch-up. The base case is that agent adoption continues to compound and infrastructure spend grows with it, allowing today's high multiples to be absorbed by earnings growth rather than multiple compression. The upside case is Bernhardsson's tenfold-developer world: if AI agents become the default development workflow, the addressable market for code-execution infrastructure expands far beyond today's AI-native companies to every engineering organization on earth. The downside case is the 700-firm study writ large: output rises, shipped software does not, and chief information officers pause agent spending until the productivity claim is proven in shipped features rather than generated lines.
In the long term, the structural question resolves around who owns the primitive. If code execution becomes a standardized utility, margins compress and value migrates to the model layer or the application layer. If it remains a differentiated, ergonomics-driven platform, the infrastructure specialists can defend premium pricing. The companies most exposed on the wrong side of that split are traditional cloud providers that treat agent workloads as ordinary virtual-machine demand, and coding-agent startups whose only advantage is a user interface rather than a proprietary execution layer. Between those two, the interface-only startups are the more fragile: a model laboratory can add a coding agent as a feature, but a coding-agent startup cannot easily add a frontier model.
For investors watching the space, the practical signals are narrow and observable. First, the AI-authored share of merged code: sustained growth above 22 percent supports the structural read; stagnation supports the cyclical one. Second, review-cycle time: if the 49 percent increase in pull-request review duration persists while output stays flat, the review bottleneck is the binding constraint, not compute. Third, customer concentration: if the largest AI laboratories begin announcing in-house execution platforms, the vertical-integration risk is no longer theoretical. And fourth, unit economics: if revenue grows faster than compute costs, the operating-leverage story is intact; if compute costs grow faster, the category is subsidizing its own growth.
The investors negotiating Modal's next round are effectively betting that the bottleneck of software creation has moved, permanently, from writing code to running it. The rest of the market now has to decide whether to follow.
The takeaway: Modal's $15 billion aspiration is not a bet on better models; it is a bet that the scarce resource in AI software is no longer intelligence, but trustworthy, instant code execution. If that is right, the infrastructure layer will capture the decade. If the 700-firm study is right instead, the market is paying for lines of code that never ship.
Explore more exclusive insights at nextfin.ai.
