409A Valuations for AI / ML Startups

How this page fits with our other guides: This article covers what is specific to AI and machine learning companies. The general mechanics of valuing a company before it has revenue are in our guide to 409A valuation methods for pre-revenue startups, and we link to it rather than repeat it here.

Why AI Startups Are Different for a 409A

The rules are the same for every company. Treasury Regulations Section 1.409A-1(b)(5)(iv) requires a reasonable valuation method that considers all available information material to value, and the result sets the lowest exercise price at which options stay exempt from Section 409A. What changes for AI companies is the evidence the appraiser has to work with. Five patterns come up again and again:

• Capital before revenue. Training and serving models is expensive, so AI companies often raise large seed and Series A rounds before they have a product in market. The financing, not the income statement, carries most of the valuation evidence.
• Fast-moving preferred prices. In a hot category, the price per preferred share can step up sharply in a year. Each step-up is material information for the 409A.
• Unusual cost structure. Inference and training costs scale with usage in a way traditional software hosting does not, which compresses gross margin and makes revenue multiples harder to borrow from SaaS.
• Thin, volatile comparables. The public AI peer group is small, concentrated and more volatile than broad software indexes.
• Talent-heavy equity. Researchers and engineers often receive large grants, so the common stock value directly drives hiring cost and retention. The stakes of the number are high.

None of these patterns calls for a special AI method. They call for careful use of the standard methods and clear documentation of the judgments the appraiser makes.

How is a 409A valuation done for an AI startup?

A 409A valuation for an AI startup has three steps. The appraiser values the whole company, usually by backsolving from its latest preferred round, sometimes cross-checked with market comparables. That value is then allocated across share classes. Finally, a discount for lack of marketability is applied to reach the fair market value of one common share.

In practice, the approach depends on how much operating history the company has. The table below shows how an AI company 409A valuation typically shifts across the main business models in the sector.

For a deeper look at when each approach applies, see our guide to the income, market and cost approaches in a 409A.

The Mega-Round Problem: Large Raises With Little Revenue

When a company has a recent arm's-length financing, the price investors paid is usually the best evidence of value available. The appraiser uses an option pricing model (OPM) to find the total equity value at which the new preferred shares are worth exactly what investors paid, then reads off the value of the common stock. This is the OPM backsolve method, and it is the workhorse of the AI startup 409A valuation.

The backsolve assumes the round reflects fair value for a cash investment. AI rounds can stretch that assumption in a few ways, and the appraiser needs to ask about each:

• Non-cash consideration. Some strategic investors in AI have funded rounds partly with cloud or GPU credits. If the credits are restricted to one provider, expire, or exceed what the company can use, they may be worth less to the company than their face value. The cash-equivalent price per share is then lower than the headline price.
• Strategic premiums. A chip maker, cloud provider or large software company may pay more than a financial investor would because it also gets a commercial relationship. That premium belongs to the strategic deal, not to the common stock.
• Unusual terms. Senior liquidation preferences, multiples above 1x, participation, or warrants issued alongside the round all shift value toward the new investor. The OPM must model those terms; ignoring them overstates common.
• Small or insider-led rounds. An extension priced by existing investors is weaker evidence than a new lead investor setting the price at arm's length.

If the round is not a clean anchor, the appraiser can adjust the price, give the backsolve less weight, or rely more on a market approach. Whatever the choice, the report should explain it. For the wider question of whether to treat a fundraise as a trigger for a new report, our guide to a 409A refresh after a funding round covers the timing.

Why is an AI startup's 409A so far below its preferred price?

An AI startup's 409A is below its preferred price because common and preferred are different securities. Preferred shares get their money back first in a sale, often carry other protections, and are bought by investors who negotiated those rights. Common has none of that and cannot be sold freely, so it is worth less per share.

A worked example shows the size of the gap. Assume an AI startup with this capital structure, and that it has just closed a Series A:

• 12.0M common shares, including options, treated as outstanding common for simplicity
• 3.0M Seed preferred shares issued at $2.00 ($6M raised)
• 6.0M Series A preferred shares issued at $10.00 ($60M raised)
• Both preferred series carry a 1x non-participating liquidation preference and rank equally
• OPM inputs: 3 years to exit, 70% volatility, 4.0% risk-free rate

The headline post-money valuation is 21.0M shares × $10.00 = $210M. The backsolve tells a different story:

Two things stand out. First, the backsolved equity value of $155.3M is well below the $210M post-money figure, because the post-money calculation prices every share as if it were Series A preferred. Second, the common stock fair market value is about half the price investors paid. A larger preference stack, senior terms, or participation would push the ratio lower. Our explainer on the 409A price vs the preferred price covers the general drivers of this gap.

For a hot AI round, the gap matters in both directions. Founders want a low strike price for recruiting, but an AI startup 409A valuation cannot ignore a large step-up just because it is inconvenient. The backsolve keeps the answer tied to what investors actually paid.

Choosing Comparable Companies for an AI Company 409A Valuation

Comparable public companies feed a 409A in two ways. They supply revenue or earnings multiples for the market approach, and they supply the volatility used in the OPM and the DLOM. For AI, both uses are harder than usual.

The public companies most associated with AI are very large, profitable and diversified, such as the major semiconductor and cloud companies. Their multiples reflect scale and earnings that a startup does not have. Smaller listed pure-play AI companies exist, but there are few of them and their trading can be erratic. A careful appraiser usually builds the set in layers:

1. Business-model peers first. An AI legal tool is first a legal software company; an AI developer platform is first a developer-tools company. Start with companies that sell to the same customers in the same way.
2. Adjust for margin and growth. If the subject's gross margin is well below the peers' because of inference costs, a straight revenue multiple overstates value. Appraisers adjust the multiple or use gross profit as the base.
3. Add AI-exposed companies for volatility. Even when they are poor sources of multiples, AI-exposed public companies help show how the market prices risk in the sector.
4. Document why each company is in or out. A set made only of the highest-multiple AI names will draw questions from an auditor.

Private transaction data, such as recent rounds at similar AI companies, can be a useful cross-check. It is noisy, though, because private round prices are preferred prices with different terms, so they cannot be compared directly with a common share value.

Volatility and the DLOM for AI Companies

Volatility is the input most affected by AI comparables. AI-exposed stocks have swung more than broad software indexes, so a peer-based estimate for an early AI company can land at the high end of the range appraisers use for startups. Our guide to Black-Scholes volatility inputs explains how the figure is measured and adjusted.

Many founders expect that a higher volatility means a lower 409A. In a backsolve, that is not how it works. Re-running the example above at different volatilities shows why:

Higher volatility gives common, which behaves like a call option on the company, more value before the discount. It also raises the DLOM, because a put option protecting an illiquid share is worth more when the share is more volatile. In this example, the two effects almost cancel. The result depends on the capital structure and the DLOM model, but the lesson holds: volatility is not a dial for lowering the strike price. Our comparison of DLOM models shows how the choice of model changes the discount.

Compute Costs, Gross Margin and the Income Approach

Compute is the line item that most separates an AI company from a traditional software company, and it shows up in the 409A in three places.

• Gross margin. Serving a model costs money every time a customer uses it. AI applications often report gross margins below the levels typical of mature SaaS. Appraisers look at the margin trend: falling model prices, smaller models and better caching can improve it, while heavier usage on flat-price plans can make it worse.
• Runway. Training runs and committed cloud contracts consume cash quickly. Runway affects the time-to-exit input in the OPM and whether a down-side scenario, such as a sale below the preference stack, belongs in the analysis. For companies with a short runway, a PWERM scenario analysis can capture those outcomes better than an OPM alone.
• Commitments. Multi-year, take-or-pay compute contracts are obligations that may not appear as debt on the balance sheet. The appraiser should know about them because they affect cash needs and the risk of the business.

The income approach, a discounted cash flow, is rarely the main method for an early AI company. Forecasts depend on model costs and prices that can move sharply in a year, so they are hard to support. Once an AI company has steady revenue and a reliable forecast, a DCF can add useful weight, especially as a check on what the market approach implies. Compute assumptions are then among the first things reviewers test.

Valuing Model Moats, Proprietary Data and IP

Founders often ask whether their model weights, training data or research team are valued separately in the 409A. Generally they are not. A 409A values the equity of the business as a whole. Intangible assets matter because they affect the inputs: how durable revenue is, what multiple is reasonable, and how much risk the forecast carries.

That is where the moat question comes in. Appraisers and investors ask the same things about an AI company:

• Can a better general model replace it? A thin layer on top of a third-party model is exposed every time that model improves. That risk supports a lower multiple or a higher discount rate.
• Is there proprietary data? Exclusive data, workflow integration and customer feedback loops are harder to copy than a prompt.
• Who owns the IP? Clean assignment of inventions from founders and contractors, and clear rights to training data, reduce legal risk that could otherwise weigh on value.

The cost approach, which values a company by what it would cost to recreate its assets, sometimes comes up for very early AI companies. It rarely captures the value of a team or a model in progress, so it is usually a floor rather than the main indication.

Exit patterns in AI also deserve attention. Some AI companies have been absorbed through licensing-and-hiring arrangements rather than a conventional acquisition, in which investors may recover their money while common holders receive little. If that kind of outcome is realistic for the company, a scenario analysis should reflect it.

How often does an AI startup need a new 409A?

An AI startup needs a new 409A at least every 12 months and sooner whenever something material happens. For fast-moving AI companies, that often means after every priced round, a large tender offer, a major customer or partnership win, or a sharp change in compute costs. Many AI companies end up updating every six months or more often.

The 12-month rule comes from the safe harbor in Treasury Regulations Section 1.409A-1(b)(5)(iv)(B)(2)(i): a valuation by a qualified independent appraiser as of a date no more than 12 months before the grant is presumed reasonable, and the IRS can overcome that presumption only by showing the valuation is grossly unreasonable. The same regulations say a valuation is not reasonable if it fails to reflect information available after its date that may materially affect value. In a sector where prices move quickly, the material-information test usually bites first.

Common triggers for an AI company include:

• A new priced round or a large extension, especially at a higher price
• A tender offer or significant secondary sale of common stock
• A strategic partnership with an investor or a large compute provider
• A product launch or model release that changes revenue expectations
• Acquisition interest or a signed letter of intent
• A drop in a key customer, a pricing collapse, or a failed financing

Granting options on a stale 409A is the most common way AI companies get into trouble. If options are granted below fair market value and no exemption applies, they can be treated as deferred compensation under Section 409A. If they then fail to comply, the holder may owe income tax on vested amounts, an additional 20% federal tax and premium interest, and some states, such as California, add their own tax. See our 409A safe harbor guide for the full rules.

What AI Founders Should Prepare for the Appraiser

A 409A for an AI startup goes faster and holds up better when the appraiser has the AI-specific facts up front. Beyond the usual cap table, charter and financials, prepare:

• Full terms of the latest round, including any credits, warrants, side letters or commercial agreements signed with investors
• Compute spend by month, split between training and inference, and any committed cloud contracts
• Gross margin by product, with the trend over recent quarters
• Which third-party models the product depends on and what it would take to switch
• Details of any secondary sales or tender offers, including buyer, price and volume
• Current runway and the plan for the next financing

Companies with an established SaaS business and AI features can also use our guide to 409A valuations for SaaS companies, since many of the same revenue metrics apply.

The Bottom Line on 409A Valuations for AI Startups

A 409A valuation for an AI startup follows the same rules as any other, but the evidence looks different: large rounds before revenue, fast-moving preferred prices, volatile comparables and a cost structure shaped by compute. The backsolve is usually the anchor, and the appraiser's main job is to test whether the latest round is a clean anchor, choose comparables that fit the business model, and capture compute and model risk in the inputs.

For AI and machine learning founders, the practical points are simple. Expect common to sit well below the preferred price. Do not expect a higher volatility to lower it. Update the valuation whenever something material happens, which in AI is often. And give the appraiser the full terms of every financing, because non-cash and strategic terms are where AI rounds most often differ from the headline price.

The figures in this article are illustrative, based on simplified inputs (options are treated as outstanding common, and a single exit horizon is used), and are not a valuation of any company. This article is general information about Section 409A and valuation practice, not legal, tax or accounting advice. Whether a valuation qualifies for a presumption of reasonableness under Treasury Regulations Section 1.409A-1(b)(5)(iv)(B) depends on the specific facts; consult your own counsel and tax advisor.