Is the GaaS Valuation Bubble Real? Inside the Agentic AI Funding Debate

Table of Contents

• What People Actually Mean by "Bubble"
• The Numbers Behind the Argument
• The Bull Case: Why the Multiples Might Be Earned
• The Bear Case: Where the Math Breaks
• The Revenue-Quality Problem at the Center of It All
• Why GaaS Bubbles Differently Than SaaS Did
• What a Correction Would Actually Look Like
• Insights Most People Overlook
• References

What People Actually Mean by "Bubble"

The word gets thrown around so loosely that it has almost stopped meaning anything. When a partner at a tier-one fund says "this isn't a bubble" and a public-market short-seller says "this is the most obvious bubble since 1999," they are frequently not even disagreeing, they are talking about different things.

There are at least three separate claims hiding inside "GaaS valuation bubble":

1. Prices are detached from fundamentals. Multiples have run far ahead of revenue, margins, and retention.
2. The capital is mispriced relative to risk. Investors are underwriting agent companies as if they were durable software businesses when the risk profile is closer to a consulting firm with a thin technology moat.
3. The whole category is built on a temporary arbitrage, namely, that today's frontier models are good enough to demo but not yet good enough to reliably do the work being sold, and the gap is being papered over with human-in-the-loop labor and aggressive marketing.

You can believe any one of these without believing the others. A company can be wildly overpriced (claim 1) yet sit on genuinely durable revenue (claim 3 fails). Another can have beautiful retention yet still be a bad investment because you paid 90x for it (claim 1 true, claim 2 true, claim 3 false). The debate only gets useful once you separate these threads, because the policy implications differ. If it's just price, you wait for a reset and the good companies survive. If it's revenue quality or model dependence, even the "winners" can disappoint.

The Numbers Behind the Argument

Here is what is not in dispute. Agentic AI has pulled in an extraordinary share of venture dollars. By the back half of the 2020s, AI broadly was absorbing a large plurality of all US venture funding, and within that, application-layer agent companies were commanding the kind of round sizes and valuations that used to be reserved for late-stage growth equity, except now they show up at seed and Series A.

Several patterns recur:

• Seed rounds at nine-figure valuations for teams of fewer than ten people, sometimes pre-revenue, often pre-product. This is the data point bears point to first, and it is genuinely hard to defend on classic terms.
• Forward-revenue multiples in the 30x to 100x range, versus the 10x-to-15x that even premium SaaS commanded at the peak of the 2021 cycle. (For context on how unusual that is, the long-run median for public software has historically sat in the single digits to low teens, per Bessemer's State of the Cloud analysis.)
• Revenue that is real but young, many of these companies genuinely are adding ARR faster than any prior software cohort. The fastest are reportedly hitting $1M, $10M, even $100M in revenue in timeframes that would have been considered fictional five years ago.

That last point is what makes the debate hard. You cannot simply say "no revenue, therefore bubble." The revenue is there. The fight is about what it's worth and how long it lasts. McKinsey's work on the economic potential of generative AI estimates trillions in annual value creation, which the bulls cite as the denominator that makes today's valuations look small; the bears note that "value created across the economy" and "value captured by the agent vendor you funded at 80x" are very different numbers. You can read McKinsey's framing in their report on the economic potential of generative AI.

The Bull Case: Why the Multiples Might Be Earned

The strongest version of the bull case is not "AI is magic." It's an argument about market size and substitution.

Traditional SaaS sold you a tool and charged per seat. The buyer still had to hire the humans to use the tool. Agentic AI, sold as a service, proposes to sell the outcome, to replace or compress the labor itself. That changes the addressable market from "software budgets" to "labor budgets," and labor budgets are roughly an order of magnitude larger. If even a sliver of services spend converts to agent spend, the revenue ceiling for the category is enormous. This is the core of why agent startups command premium valuations in the first place, and it's a genuinely different thesis than the seat-based one VCs spent two decades pattern-matching on.

The bull also points to growth velocity as a leading indicator of product-market fit. When a company adds $50M of net new revenue in a year with a small team, that is not nothing, even if you think the multiple is rich, the slope of the curve tells you customers are getting value today, not in some speculative future. Andreessen Horowitz has argued repeatedly that the application layer, not the model layer, is where durable enterprise value accrues, on the logic that workflow lock-in and proprietary data loops compound over time; their broader thinking on this lives in their writing on AI and the enterprise.

And finally: capital efficiency is improving, not worsening, for a subset of these companies. The best agent startups are reaching meaningful revenue with tiny headcounts because the product is the labor. A ten-person company doing $40M in revenue is a different animal than a 400-person SaaS company doing the same, and if model costs keep falling, the margin story improves with time rather than degrading.

The Bear Case: Where the Math Breaks

The bears have three good arguments and one great one.

Argument one: the cost of goods sold is unlike anything software has seen. Every agent action burns tokens, and complex autonomous workflows can burn a lot of them, chains of reasoning, tool calls, retries, verification passes. This is the burn-rate problem that is specific to agents: unlike SaaS, where serving the millionth customer costs almost nothing, serving the millionth agent task can cost real money. Gross margins that look like 80% software margins on a pitch deck can be 40% or 50% in practice once inference is fully loaded. Price a company on SaaS margins when it has consulting margins and you have mispriced it by design.

Argument two: model dependence is an existential moat problem. Many application-layer agents are, structurally, a prompt-and-glue layer on top of a foundation model they don't control. The lab that supplies the model can move up the stack and offer the same workflow natively, and has every incentive to. The "agentwashing" problem in fundraising decks is a symptom: when everyone claims a proprietary agentic moat, the ones that actually have one become hard to distinguish from the ones reselling someone else's intelligence.

Argument three: enterprise pilots are not enterprise revenue. A huge share of the headline ARR is pilot and proof-of-concept money, which renews at a fraction of the rate of committed SaaS contracts. Pilots convert to production far less often than the growth charts imply, and when budgets tighten, pilots are the first line item cut.

The great argument ties them together: you are paying SaaS-bubble multiples for businesses whose revenue is less durable, whose margins are thinner, and whose moats are rented. That's not one risk stacked on a strong base, it's three risks that all point the same direction. As HBR has noted in its analysis of why most enterprise AI initiatives stall before delivering returns, the gap between the AI hype and measurable enterprise ROI remains wide, and the buyer eventually notices.

The Revenue-Quality Problem at the Center of It All

Strip away the noise and almost every serious disagreement reduces to one question: is GaaS revenue recurring, or is it consumption that's been styled to look recurring?

This matters because valuation multiples are, at heart, a bet on durability. You pay 50x forward revenue because you believe that revenue will still be there, and growing, in five years. Net revenue retention is the metric that proves it. A SaaS company with 130% NRR earns its multiple because the existing book of business expands on its own. The trouble with much GaaS revenue is that it's usage-based and outcome-tied, which cuts both ways: in a good year, a customer who uses the agent more pays more, and NRR looks spectacular. In a bad year, a recession, a budget freeze, a better model from a competitor, usage can fall off a cliff in a single quarter, something seat-based SaaS almost never experiences.

So the same revenue line can be the bull's best evidence and the bear's biggest fear. Usage revenue is higher-beta. It amplifies the cycle in both directions. The companies that will survive a valuation reset are the ones converting usage into commitment, minimum spend floors, annual contracts, embedded workflows that are painful to rip out. The ones that won't are riding pure consumption with no switching cost, where the customer can turn the agent off as easily as they turned it on. Underwriting that difference is now the entire job of a late-stage GaaS investor, and most term sheets are not yet doing it well.

Why GaaS Bubbles Differently Than SaaS Did

It's tempting to reach for the 1999 or 2021 playbook, but the analogy is imperfect in ways that matter for how this resolves.

In the dot-com bust, many companies had no revenue and no path to it; the correction was brutal and total. In the 2021 SaaS reset, the businesses were mostly real, the problem was purely the multiple, so prices fell 60-70% while the companies kept compounding and eventually grew back into reasonable valuations. GaaS sits in an uncomfortable third spot: the revenue is real (unlike 1999) but its quality and cost structure are genuinely in question (unlike 2021). That means a GaaS correction is likely to be more discriminating than either prior one.

A pure multiple compression, the 2021 scenario, punishes everyone equally and rewards patience. A revenue-quality reckoning, the scenario unique to agents, sorts the field. It separates the companies sitting on durable, committed, defensible revenue from the ones whose ARR evaporates the moment a foundation lab ships the feature natively or the customer's pilot budget dries up. The first group survives a haircut and thanks you for the buying opportunity. The second group doesn't get a down round; it gets a shutdown.

This is also why "is there a bubble?" is the wrong question for an operator or investor. The useful question is which layer and which revenue type is overpriced. Infrastructure and "picks and shovels" plays bubble differently than vertical application agents. Outcome-priced agents in regulated, sticky verticals bubble differently than horizontal productivity agents competing directly with the model labs.

What a Correction Would Actually Look Like

If and when the reset comes, and history suggests something will, even in a category with a real future, it probably won't announce itself as a crash. It'll show up as a slowdown in the signals: round sizes flattening, the time between rounds stretching, bridge rounds quietly replacing priced ups, and a sudden vogue for the word "profitability" in pitch decks that previously sold pure growth.

The top of the market, the handful of category-defining agent companies with genuine scale, defensible data loops, and committed revenue, will likely be fine, the way the best SaaS names were fine after 2021. The graveyard fills from the middle: the well-funded Series B companies raised at peak multiples on pilot revenue that never converted, now facing a down round they can't stomach and a burn rate they can't sustain. Expect acqui-hires, fire-sale roll-ups, and a wave of post-mortems blaming "the market" for what was, in many cases, a revenue-quality problem visible in the original term sheet.

None of that invalidates the category. The internet was a real revolution and pets.com was a real catastrophe. Both things were true in 1999, and the equivalent will be true here. The job isn't to decide whether GaaS is a bubble. It's to be on the right side of the sorting when the consumption revenue gets repriced as exactly what it is.

Insights Most People Overlook

1. "Bubble" and "great category" are not opposites, they're frequently the same thing. Every genuine platform shift overshoots. The presence of a bubble is weak evidence for the category's importance, not against it. Capital floods in precisely because the long-term prize is real; the overshoot is a feature of how markets price uncertainty, not proof the thesis is wrong. The mistake is treating "it's a bubble" as a conclusion rather than a warning about timing and selection.

2. The most dangerous valuations aren't the famous nine-figure seeds, they're the quiet Series B's. The headline-grabbing mega-rounds for top teams will mostly survive because those companies have real scale and the brand to raise again. The actual carnage concentrates in the unglamorous middle: companies that raised $40-80M at peak multiples on pilot revenue, are now too expensive to acquire and too unproven to IPO, and have no cheap path to the next round. That's where the down-rounds and shutdowns cluster.

3. Falling model costs are a double-edged sword that bulls under-weight. Cheaper inference improves agent margins, good. But it also collapses the cost of building a competing agent, and it makes it trivial for the foundation labs to offer the same workflow natively. The same force that fixes your COGS dissolves your moat. Capital efficiency and defensibility move in opposite directions as compute gets cheap.

4. Outcome-based pricing is a valuation trap disguised as a revenue innovation. Charging per outcome sounds investor-friendly because it aligns price with value. But it makes revenue maximally cyclical and maximally legible to the customer, who can now see exactly what the agent costs versus what it saves, and cut it instantly when the math wobbles. Per-seat SaaS revenue was sticky partly because it was hard to attribute. Outcome pricing trades that stickiness for transparency, and transparency is the enemy of durable multiples in a downturn.

5. The best tell isn't the multiple, it's how the company answers "what happens when the model gets 10x better and 10x cheaper?" If that scenario helps them (more value to capture, deeper workflows, better margins), they're on the right side of the platform shift. If it threatens them (the lab eats their lunch, the moat evaporates), they're an arbitrage play wearing a software costume, and no multiple is safe.

References

• Bessemer Venture Partners, State of the Cloud
• McKinsey, The economic potential of generative AI: The next productivity frontier
• Andreessen Horowitz, AI writing and enterprise analysis
• Harvard Business Review, How Companies Can Actually Capture Value from AI