How AI Startups Get Funded

For a period, saying you were an AI company was itself the pitch. Capital was abundant, the category was new, and the questions were about capability.

That period ended. Investors have now funded enough AI companies to know which ones work, and the diligence has become specific and uncomfortable: what is your gross margin, what happens when the next model release absorbs your feature, and is that revenue a contract or an experiment?

This guide covers what investors actually underwrite in AI companies, the three funding profiles the sector splits into, the compute problem, and the revenue-quality questions that decide rounds.

1. Gross Margin Is the First Question Now

Classic software carries very high gross margins because serving one more customer costs almost nothing. That assumption underpins essentially every SaaS valuation multiple.

AI products built on inference break it. Every query costs money — tokens to a model provider, or GPU time on your own infrastructure — and that cost scales directly with usage. Companies building on third-party model APIs frequently find gross margins well below software norms, sometimes closer to a services business.

The consequences run through everything:

• CAC payback is longer than the same revenue would imply in software, because payback is calculated on gross profit, not revenue — see our guide to the metrics investors underwrite.
• Your most engaged users cost you the most. Under flat-rate pricing, heavy usage can be actively unprofitable — an inversion of the software instinct that engagement is unambiguously good.
• The valuation multiple compresses. Investors apply software multiples to software margins. A 45% gross margin business is not valued like an 85% one, whatever the growth rate.

What investors want to see: unit cost per query or per task tracked over time, a clear trajectory of improvement, and a specific plan — model distillation, caching, routing cheap queries to smaller models, moving to owned infrastructure at scale, or pricing that passes cost through. Founders who present this proactively are far more credible than those who omit the line and get asked in week three.

2. The Defensibility Question Has Moved

Model capability was briefly a moat. It is not one now, because capability is a fast-depreciating asset — whatever is hard today is a feature in a foundation model release within a year or two.

The "thin wrapper" critique lands whenever a product is a prompt over someone else's model with a login page. It is frequently unfair, but you must answer it directly.

What investors now accept as durable:

• Distribution. Being where the work already happens. A company that owns the customer relationship can swap models underneath and lose nothing.
• Proprietary data that genuinely cannot be reconstructed — accumulated from operations, exclusively licensed, or generated by your own users in a compounding loop. Data that could be bought or scraped is not a moat.
• Workflow depth. Being embedded in a process, integrated with systems of record, holding state and history. Switching cost, not model cost.
• Regulatory position. Certifications, audit trails, jurisdictional compliance — real barriers in healthcare, finance and government.
• Outcome accountability. Companies selling a completed outcome rather than a tool are harder to displace, because the customer has offloaded responsibility.
• Evaluation infrastructure. A rigorous internal eval suite tuned to your domain is a genuine and underrated asset, because it is what lets you adopt each new model quickly and safely.

3. Three Funding Profiles

Foundation model companies

Enormous capital requirements, driven by compute and talent. Funded by the largest growth investors, sovereign vehicles, and strategic investors — frequently with compute commitments forming part of the deal. A small number of companies, a category most founders are not in, and one where the capital requirements make conventional venture returns difficult.

Infrastructure and tooling

Orchestration, evaluation, observability, vector and retrieval infrastructure, inference optimisation, data pipelines, safety and governance tooling. Conventional venture profiles and conventional software margins, with the specific risk that foundation model providers absorb your category into their own platform.

Investors underwrite this like developer infrastructure: adoption, expansion, and whether you sit above or below the layer that commoditises.

Applications and vertical AI

Where most companies and most funding are. Applying models to a specific industry problem — legal, healthcare, finance, logistics, construction, customer support.

Investors underwrite this like vertical SaaS with two additions: gross margin, and whether the domain knowledge embedded in the product is deep enough that a general model cannot replicate it. The strongest of these look less like software companies and more like outcome-based businesses in what they charge for.

4. The Compute Problem

Compute is the largest input cost for many AI companies, and it has produced financing structures that did not previously exist.

• Cloud credits and startup programmes. Meaningful non-dilutive value, particularly at early stage. They also create switching costs, which is the point of them — so understand what happens when the credits expire and model the cliff.
• Compute commitments in investment deals. Strategic investors, including cloud providers, may invest partly in the form of compute. This is real value but it is not cash: you cannot pay salaries with it, and it locks your infrastructure choices. Value it accordingly rather than counting it at face against the round size.
• Equipment financing for owned hardware. Companies running their own GPUs can finance them as assets rather than funding them from equity, which is dramatically cheaper capital for a predictable workload.
• Reserved capacity and committed spend, which trades flexibility for lower unit cost — sensible once demand is predictable, dangerous before.

Founders should be clear-eyed that credits and compute commitments flatter early margins. Investors model what happens at list price, and so should you.

5. Revenue Quality: Pilots Versus Production

This is where the most rounds are lost, and it is entirely avoidable.

Enterprises have run large numbers of AI pilots. Budget exists, procurement is willing, and a founder can accumulate an impressive-looking revenue figure from experiments that will not renew.

Investors separate these ruthlessly. Expect to be asked:

• What proportion of revenue is production deployment versus pilot or proof of concept?
• What has actually renewed, and at what rate? A first renewal cohort is the single most informative data point you have.
• Who holds the budget — an innovation budget, or the line owner whose work the product does? Innovation budgets disappear.
• Is pricing per seat, per usage, or per outcome? Usage-based pricing aligns revenue with cost, which is increasingly favoured; seat-based pricing in an agentic product raises awkward questions about what happens as the product replaces seats.
• What is net revenue retention within the cohort that got past pilot?

A company with modest but genuinely committed production revenue raises more easily than one with a larger number built from experiments. Presenting the split yourself, unprompted, is a strong credibility signal.

6. What Diligence Now Includes That It Did Not

• Model dependency. What happens if your provider changes pricing, deprecates a model, or ships your feature? Investors want to see abstraction and a tested fallback, not a plan to worry about it later.
• Evaluation rigour. How do you know the system works? Domain-specific evals, regression testing against model updates, and measured error rates. "It works well in demos" is not an answer.
• Data rights. What are you permitted to train on? Customer data used for training without clear contractual rights is a serious liability, and it surfaces in diligence and again at exit.
• Regulatory exposure. Sector rules in healthcare and finance, plus emerging horizontal AI regulation with obligations that vary by risk classification and jurisdiction. Investors want to know you have identified which apply, not that you have solved everything.
• Failure mode and liability. What happens when the system is wrong, who bears the loss, and what does your contract say? For products taking consequential actions, this is a central question and increasingly an insurance one.
• Talent concentration. Senior AI compensation is high and mobile. Investors examine retention and whether capability sits with one person.

7. Pricing Is Where the Margin Story Is Won or Lost

Gross margin is not only a cost problem. Most AI companies with poor unit economics have a pricing model inherited from software, applied to a product whose costs behave nothing like software.

Four models are in use, and investors read your choice as a statement about how well you understand your own business.

• Per seat. Familiar to buyers and easy to forecast, and increasingly awkward. If the product does work a person used to do, seat counts fall as the product succeeds — you have priced yourself against your own value. Expect to be asked how you avoid that.
• Per usage. Tokens, queries, documents, minutes. Cost and revenue move together, which protects margin, but buyers dislike unpredictable bills and procurement teams frequently refuse them outright. The common fix is a committed floor with usage billed above it, which delivers predictability and protection at once.
• Per outcome. Per resolved ticket, per approved claim, per completed reconciliation. The strongest position commercially, because the buyer is purchasing a result rather than a tool — and the hardest to operate, because you must define, measure and defend what counts as an outcome while carrying the cost of the failures.
• Hybrid. A platform fee covering access plus a usage or outcome component. This is where most serious enterprise AI pricing has landed, and it is the easiest to explain to an investor because it cleanly separates the recurring base from the variable layer.

Two pricing mistakes recur often enough to name. The first is flat unlimited pricing offered early to win logos, which converts your best customers into your largest losses and is extremely difficult to unwind later. The second is anchoring price to the incumbent software tool rather than to the labour cost of the work being replaced — an anchor that caps your realisable price at a fraction of the value you deliver.

Investors also look closely at whether you have ever raised prices. A company that has repriced an existing cohort and kept it has demonstrated something no growth chart can: that customers value the product above what they are currently paying.

8. Practical Advice for Raising

• Lead with the problem and the customer, not the technology. Every deck says AI. Very few say precisely whose job gets measurably better.
• Put gross margin on the slide. With the trajectory and the plan. Omitting it invites the worst assumption.
• Split pilot from production revenue yourself.
• Answer the wrapper question head on. Name what you would still have if the underlying model became free tomorrow. If the honest answer is nothing, that is worth knowing before an investor says it.
• Show the eval suite. It demonstrates engineering seriousness better than any architecture diagram.
• Match investor to profile. Infrastructure, application and foundation model companies need different investors with different return models — as our guide to fund economics explains.

Frequently Asked Questions

Do we need our own model?

Usually not, and training one is an expensive way to build a depreciating asset. Fine-tuning, retrieval over proprietary data, and strong orchestration deliver most of the benefit at a fraction of the cost. Own models make sense where privacy, latency, cost at very high volume, or genuinely specialised domains demand it.

How do investors think about open-weight models?

Favourably, as an option rather than an identity. Running open-weight models on your own infrastructure can transform gross margin at volume and answers data-residency objections in regulated sectors, at the cost of engineering burden and a capability gap against the frontier that varies enormously by task. The credible position is that you have benchmarked both on your own evals and route by task — not that you are committed to either on principle.

What do investors ask about agentic products specifically?

Three questions before any others: what happens when the agent is wrong, who approves consequential actions, and how cost per completed task behaves as task complexity rises. Multi-step agents can consume many times the inference of a single query, so margin measured at demo scale tells you very little. Show cost per successfully completed task, not cost per call.

Does a large waitlist or viral free usage help?

Less than founders hope, and occasionally it hurts. Free usage on an inference-cost product is a liability with a growth chart attached. Investors will ask what the free tier costs you every month and what proportion converts — and a company with a smaller number of paying production customers usually raises more easily than one with a large unmonetised audience.

How much should an AI seed round be?

Whatever gets you to evidence of production usage and a defensible margin trajectory. Rounds in this category have run larger than historical norms, partly justified by compute costs and partly by competition. Over-raising at a valuation you cannot grow into remains the same trap it always was — see our guide to down rounds.

Is AI-generated code a diligence problem?

It is now a standard question. Investors and acquirers ask about provenance, licensing and whether your invention assignment agreements cover AI-assisted work. Have a policy, and make sure your counsel has reviewed it.

Should we take strategic money from a model provider or cloud?

It brings compute, credibility and technical access. It can also signal alignment that complicates relationships with their competitors, and it may constrain your infrastructure choices. Resist rights of first refusal over your company, and value compute-in-lieu-of-cash honestly rather than at headline.

Are there non-dilutive options?

Yes, and they are under-used. R&D tax credits apply squarely to this work, cloud credit programmes are substantial, equipment financing works for owned GPUs, and applied AI in defense, health and energy is directly fundable through SBIR and STTR.

The Bottom Line

The AI premium in fundraising has narrowed to companies that can answer three questions well: what is your gross margin and where is it heading, what do you still have if the model layer commoditises, and how much of your revenue is production rather than experiment.

Answer those unprompted, with numbers, and you are ahead of most of the market.

Global Capital Network connects AI founders with investors who understand these economics. See upcoming events or get in touch.