Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstalliTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI startups do not share one economic model. Some sell reusable software with modest per-customer costs; others incur substantial, recurring expenses for model inference, cloud capacity, data operations, or human review. Investors should compare each company’s cost to serve, ability to scale, defensibility, capital needs, pricing, and retention—not assume that an AI label predicts weaker margins or better growth.
Why an AI label is not a business model
Traditional software companies are often described as building a product once and selling it repeatedly. That model can support recurring revenue and high gross margins because serving another customer may add little cost after development. But “traditional software startup” covers a wide range of businesses: implementation-heavy enterprise products, hosted software, and low-touch subscriptions do not have identical economics.
AI companies vary just as much. A product that calls a third-party model to generate a response has a different cost structure from one that trains or serves its own large model. A feature embedded in a broader software product may carry different economics from a product whose primary value is processing large volumes of customer data. Some AI products automate a complete workflow; others still depend on people to correct outputs or prepare data.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsThe relevant question is therefore not whether AI is inherently less scalable than conventional software. It is whether the company can deliver the customer’s desired outcome reliably, at a cost that leaves room for attractive margins and growth. An AI product can scale well when its workload is efficient and repeatable. A conventional software product can also have weak economics if customer-specific implementation and support absorb the revenue.
#1 Best Overall
Compare cost to serve, not just gross-margin headlines
Reusable software can have low marginal delivery costs, but AI products may add expenses that recur with use. The most visible is inference: the computing cost of generating predictions, text, images, or other outputs. Hosting and cloud infrastructure may also rise with workload. Other costs can include data labeling and maintenance, evaluation, quality control, customer-specific tuning, and human review.
These costs are not automatically all recorded in the same place in a company’s accounts. During diligence, establish what management includes in cost of revenue. If model fees, GPUs, cloud hosting, data operations, or customer support are excluded from that figure, the reported gross margin may not show the full cost of delivering the product. Compare gross margin with contribution margin and gross profit per customer or completed outcome, using a consistent definition.
A dated margin estimate is context, not a benchmark
An a16z article published around 2020 estimated gross margins of 50–60% for AI companies, compared with 60–80% or more for comparable SaaS businesses. The article’s source note says the AI estimate came from founder interviews. Treat it as an anecdotal, dated investor observation—not a representative current-sector dataset, a forecast, or a target for an individual company. A company’s architecture, pricing, workload, and accounting definitions matter more than that comparison.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Questions that expose the actual unit economics
- What is the fully loaded cost of a typical task, customer, or completed outcome, including model/API fees, compute, support, data work, and review?
- How do cost and quality change as usage increases? Does a longer prompt, larger file, more demanding task, or higher service level materially change the cost?
- What share of outputs need correction, escalation, or human approval, and who pays for that work?
- Does the company earn more gross profit as a customer uses the product more, or does additional usage consume the revenue?
- Are customer-specific tuning and integrations included in implementation revenue, ongoing support, or an unpriced cost of sale?
Look at trends by customer cohort and workload, not only a company-wide margin. If margins improve as the product is used more and customers mature, that can indicate an improving delivery system. If demanding customers require bespoke model work or persistent review, aggregate growth may hide a cost-to-serve problem.
Test whether delivery scales beyond the first set of customers
Once conventional software has been built, adding another customer can require little additional labor when setup and support are standardized. AI products can face a different operational burden: unusual cases, inconsistent input data, evaluation of output quality, integration into existing systems, and decisions about when a person must intervene. These tasks may be essential to a trustworthy product, but they affect how quickly revenue can grow without a matching increase in labor.
The key distinction is not “automated” versus “not automated” in a product demo. It is what happens in production across varied customer workflows. A system that completes an end-to-end task reliably may scale efficiently. A system that generates a first draft but leaves customers or company staff to verify and repair it may deliver less automation than its headline feature suggests.
Measure the work hidden behind deployment
- Review and correction: Track the portion of outputs that require human review, correction, or escalation, and whether that portion changes as the system improves.
- Onboarding: Separate repeatable setup from work that requires one-off data cleanup, process mapping, or custom engineering.
- Implementation burden: Compare time and labor per deployment across customer cohorts. Determine whether the company can serve the next customer using the same product or needs a new engineering branch.
- Reliability: Examine performance on edge cases and the process for monitoring quality after launch, not only results on a narrow demonstration task.
- Customer ownership: Find out who carries the operational burden when an output is wrong or incomplete, and whether that burden affects renewal or expansion.
These measures help distinguish software economics from service economics. Services can be valuable and profitable, but if each new contract requires a similar amount of expert labor, revenue may not scale like a reusable software product. A company can improve this profile through better workflows, tools, or automation, but investors should evaluate observed progress rather than assume it.
Look for defensibility in workflow, context, and distribution
Access to a capable model can enable a product, but access alone may not protect it. If competitors can use broadly available models to reproduce a feature, the model itself may not create a durable advantage. An a16z analysis cautioned against treating model access as a moat; PwC’s M&A commentary points to domain depth, ecosystem entrenchment, proprietary context, and mission-critical workflow integration as potential sources of resilience. These are useful diligence hypotheses, not guarantees.
Investigate what the company has that a competitor cannot quickly replicate. It may be a trusted position in a customer workflow, specialized domain expertise, useful integrations, an established route to customers, or customer-specific context that improves the product. The value of any data advantage depends on whether the company can lawfully and contractually retain and use the data, and whether that use actually improves results.
Questions for evaluating a claimed moat
- What customer context or data can the company retain and use under its contracts and applicable rights?
- Does product usage improve quality or performance in a way that competitors cannot readily copy, or is the claimed learning effect still unproven?
- Is the product embedded in a workflow customers depend on, with meaningful switching costs or integration depth?
- Could an incumbent software vendor or model provider add the same feature, and what would still make customers choose this company?
- Is distribution, customer trust, or domain knowledge more important to winning and retaining business than model differentiation?
Do not equate a large volume of data with defensibility. Data may be difficult to use, weakly connected to product quality, or unavailable for reuse. Likewise, workflow integration can create resilience but does not guarantee a moat if an alternative is easy to adopt or the product’s value is limited. Test the mechanism behind each claimed advantage.
Separate capital intensity from the AI category’s funding headlines
AI funding totals show where investor capital is flowing; they do not establish startup-level unit economics, likely returns, or the amount any particular company needs. The OECD’s 2026 analysis of 2025 venture-capital data reports that AI firms received USD 258.7 billion, or 61% of USD 427.1 billion in global VC investment. About 73% of AI investment value came through mega-deals over USD 100 million. That concentration matters: a headline share dominated by large deals does not mean funding was distributed evenly among startups.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Infrastructure also absorbed substantial capital. The OECD reported USD 109.3 billion in 2025 VC investment for AI IT infrastructure and hosting companies, and cumulative investment of USD 256.1 billion in that category from 2012 through 2025. Those are capital-flow figures for a category, not evidence that an application company must spend comparable sums or that infrastructure investment guarantees returns.
Rank #4
- Author: Guillebeau, Chris.
- Publisher: Currency
- Pages: 304
- Publication Date: 2012-05-08
- Edition: NO-VALUE
For an investor, distinguish an application startup using third-party models from a company developing, training, and serving large models or building infrastructure. Their compute requirements, assets, supplier exposure, and funding profiles can differ sharply. A software application may have material usage expenses without requiring infrastructure-scale investment. Conversely, a company promising model development or large-scale hosting may need substantial capital before its economics can be tested.
Use the UK survey as a scoped signal
The UK Department for Science, Innovation and Technology’s 2025 AI sector study estimated UK AI-sector revenue at £23.9 billion and employment at 86,139 for 2024; the estimates were produced by Perspective Economics. These are UK sector estimates, not a global market measure or a forecast for an individual company.
In the study’s survey, 66% of respondents (n=196) expected their need for computing power to increase, and 71% expected increased needs for skills and software or development tools. The study also found that 53% said access to equity investment had significantly affected their ability to meet business goals in the preceding 12 months. These are survey responses, not measurements of every UK AI company. The study records concerns among some firms about dependence on foreign cloud and LLM infrastructure.
Translate capital needs into a company-specific plan
- Estimate the runway required to reach the next verifiable product or commercial milestone, and identify what evidence that milestone should produce.
- Model how gross profit changes if cloud or model prices change, and ask whether the company can shift providers or workloads without damaging quality.
- Determine whether smaller or open models, caching, workload routing, or other architecture choices are feasible for the actual use case rather than merely theoretical cost-saving options.
- Separate investment in reusable capabilities from spending that simply postpones an economics test. Ask what asset or measurable improvement each major spend is expected to create.
- Assess supplier dependency: what happens to service, price, and customer commitments if a model or cloud provider changes terms or becomes unavailable?
Examine pricing and retention in the way customers buy
Seat counts may become less informative when a product automates work that previously required several users or when customers pay based on usage or outcomes. PwC’s advisory commentary argues that seat contraction can obscure performance in net revenue retention and that cohort and gross revenue retention analysis may be useful. It also notes movement toward value-based or outcome-based pricing. These observations are not a substitute for checking the company’s own contracts, renewals, usage, and evidence of customer value.
Best Value
Compare annual contract value with actual usage and cost to serve. Review retention by cohort, customer concentration, expansion and contraction, implementation expense, and pricing power. For outcome-based pricing, verify that the outcome is measurable, attributable to the product, and deliverable at a cost that supports the price. A product may create real customer value without capturing much of it in revenue; the contract and renewal data show whether that value supports a durable business.
Do not apply a conventional SaaS revenue multiple mechanically to an AI company—or dismiss one solely because it has usage-linked costs. Growth quality, gross profit, retention, supplier dependency, and the durability of the product’s advantage all shape what revenue is worth. The same revenue growth can imply different prospects when one company’s costs fall with scale and another’s rise with usage or customer-specific labor.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A practical investor comparison framework
Use the same questions across AI-native and conventional software companies, then tailor the evidence to their delivery models. The framework below is intended to structure diligence, not to label an entire category as attractive or unattractive.
Free tools Windows power users keep installed
One-click scans. No signup required.
| Comparison axis | What to establish | Warning sign to investigate |
|---|---|---|
| Revenue and cost to serve | Recurring revenue, usage patterns, fully loaded delivery cost, gross and contribution margin, and gross profit per customer or outcome. | Reported margins omit material inference, hosting, data operations, review, or customer-specific work. |
| Scale and delivery | Review and correction rates, onboarding effort, implementation labor, reliability, and repeatability across customers. | Every deployment requires bespoke engineering or a persistent layer of unpriced human work. |
| Defensibility | Workflow integration, lawful and usable context, domain expertise, distribution, customer trust, and switching friction. | The product’s only claimed advantage is access to a model that rivals can also use. |
| Capital and dependencies | Runway to the next milestone, compute requirements, supplier exposure, and the value created by infrastructure spending. | Large future spending is assumed without a clear milestone or a credible response to supplier cost changes. |
| Pricing and retention | Contract value, usage, cohort retention, customer concentration, pricing power, and evidence of realized customer value. | Seat metrics or top-line growth are treated as sufficient despite contraction, weak renewal evidence, or uncertain outcome attribution. |
Diligence checklist: what to request and test
Investors can turn the comparison into an evidence request rather than a debate over labels. Ask management to connect product usage to workload costs, customer outcomes, renewals, and the resources required to serve the next cohort.
- Map one representative customer journey. Trace a customer’s use from onboarding through a completed task. Identify software, model calls, cloud usage, data work, human review, and support at each stage.
- Reconcile reported margins. Ask for the cost-of-revenue definition and reconcile model/API fees, hosting, data operations, and review labor to the reported figure. Use consistent cohort periods and cost categories.
- Compare customer cohorts. Examine whether implementation time, correction rates, retention, and gross profit are improving as the company gains experience and product usage expands.
- Stress-test usage economics. Evaluate heavier use, more complex inputs, and changing supplier prices. Check whether pricing adjusts with workload and whether customers still receive acceptable value.
- Verify automation claims. Determine what fraction of the customer’s workflow is completed reliably, what still requires human judgment, and whether that work sits with the startup or the customer.
- Substantiate defensibility. Tie each claimed advantage to a mechanism—such as a workflow relationship, usable customer context, distribution, or domain capability—and to evidence that customers value it.
- Connect funding to milestones. Review the amount of capital needed to reach the next commercial or technical proof point, and what evidence would change the investment case if the milestone is missed.
- Test retention and realized value. Review renewals, expansion and contraction by cohort, customer concentration, and outcome measures. For outcome pricing, confirm attribution and delivery costs.
Make the comparison at company level
AI can add variable compute, ongoing data work, and review costs to otherwise reusable software; those costs can affect margins, scaling, and capital requirements. They do not make every AI business less scalable or less profitable. The investment case rests on whether a particular company can serve customers efficiently, preserve value as the product scales, and fund the path to durable demand.
Compare unit economics and operational evidence before applying category assumptions. Revenue growth, a large funding market, or a compelling model demo cannot substitute for understanding cost per task, retention, customer value, supplier exposure, and the capital needed to reach the next milestone.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
