Why Most AI Startups Are Bad Businesses - YouTube
 |
| Most AI LLM services are loss leaders. Bottom line shows Anthropic Claude is the only LLM that is currently profitable, and by a large ammount. Interactive Graphic |
Unsustainable Economics Threaten Industry
Razor-thin margins and soaring compute costs challenge viability of generative AI companies, with even industry leaders operating at substantial losses
SAN FRANCISCO — The artificial intelligence boom that has captivated Silicon Valley and attracted billions in venture capital is confronting a harsh economic reality: the vast majority of AI software companies cannot generate profits under current business models, raising fundamental questions about the industry's long-term viability.
The stark assessment comes from both industry veterans and financial disclosures. Daria Kulikova, a senior full-stack product manager with eight years of experience building B2B and B2B2C SaaS products—including a LegalTech platform generating over $500 million in annual recurring revenue—has emerged as a prominent voice challenging the AI industry's sustainability. Through her YouTube channel analyzing technology trends, Kulikova has documented how AI-native companies operate under fundamentally different economics than traditional software businesses.
While traditional software-as-a-service companies routinely achieve gross margins between 70% and 90%, generative AI startups are struggling with margins of just 30% to 60% at best, according to multiple industry reports and financial disclosures. Even mature AI products operate far below traditional software economics, with AI-powered software businesses showing 55% gross margins compared to traditional SaaS companies with 85% margins, leaving far less room for pricing errors and operating leverage.
The margin crisis extends to the industry's most prominent players. Bessemer Venture Partners' 2025 data shows fast-growing AI "Supernovas" averaging only 25% gross margins, while steadier companies trend closer to 60%, with many AI Supernovas showing negative gross margins — a phenomenon rarely seen in software.
Experience From the Trenches
Kulikova's analysis draws on her track record leading four end-to-end SaaS products and over 50 feature developments, including a B2B digital accessibility platform that scaled its user base fivefold within a year. Her perspective contrasts sharply with the venture-fueled optimism dominating AI industry discourse.
"Traditional SaaS after initial R&D and platform investment—serving more customers adds very little extra cost," Kulikova explained in her analysis. "In AI-native or GPT wrapper products, there are major ongoing costs per user: API calls, compute time, licensing, sometimes per-output moderation."
The distinction proves critical. In traditional software, customer acquisition represents the primary variable cost—typically service-related expenses like customer success managers or support specialists. In AI-native products, computational costs rise exponentially with usage, forcing companies to implement usage caps to manage expenses.
Foundational Models Bleed Cash
The economics prove challenging even for companies building foundational models. OpenAI expected approximately $5 billion in losses on $3.7 billion in revenue last year, and the company reported $4.3 billion in first-half 2025 revenue but incurred $8.5 billion in expenses, yielding a $4.7 billion loss.
OpenAI now generates $10 billion in annual recurring revenue, but CEO Sam Altman stated the company should prioritize growth and investments in training and compute "for a long time," even if it delays profitability, saying the rational approach is to "be willing to run the loss for quite a while".
Anthropic, maker of the Claude AI assistant, faces similar challenges. The company's annualized revenue jumped from $1 billion to $3 billion in just five months through May 2025, and by August 2025 reached over $5 billion in run-rate revenue. However, Anthropic expects to lose $3 billion in 2025 due to how unprofitable its models are.
The Unit Economics Problem
The fundamental issue lies in AI's cost structure. In traditional SaaS, the big upfront cost is product development, with selling copies generating pure profit afterward, but in AI each unit produced comes with material costs, similar to manufacturing widgets in a factory, as COGS now matter again.
Each user interaction requires API calls, GPU compute time, and often per-output moderation—expenses that scale directly with usage. ChatGPT was at one point costing OpenAI an estimated $700,000 per day, though costs have since declined to between $100,000 and several hundred thousand dollars daily.
The problem intensifies with heavy users. OpenAI CEO Sam Altman admitted in January 2025 that the company is losing money on its $200-per-month ChatGPT Pro plan because people are using it more than expected, with some users costing more than $200 monthly to serve.
Microsoft's GitHub Copilot experienced similar challenges. The Wall Street Journal reported Microsoft's GitHub Copilot is losing an average of $20 per month per user, with some users costing as much as $80 monthly, while Copilot charges $10 per month.
Conversion Crisis
Beyond operational costs, AI companies face severe challenges converting free users to paying customers. Despite ChatGPT's massive reach, with an estimated 500-600 million monthly active users, the platform has only 15.5 million paying subscribers, representing a conversion rate of approximately 2.6%.
Kulikova, whose product management experience includes optimizing conversion funnels for enterprise platforms, characterized this as "alarmingly low" for a product positioned as a worldwide disruptor. "In product management terms, your product-market fit goes out the window or you never had it in the first place," she noted.
OpenAI revealed 20 million paying subscribers and over 500 million weekly active users, choosing to report weekly rather than monthly metrics, with weekly active users rising 100 million while site traffic remained unchanged.
Kulikova questioned the reliance on weekly metrics, suggesting they represent "vanity metrics"—impressive numbers that don't necessarily translate to sustainable revenue. "Monthly is a much more common SaaS metric than weekly," she observed. "Why weekly? Does it really translate to revenue?"
Pricing Wars Intensify Pressure
The sector faces mounting pressure from aggressive price competition. OpenAI's GPT-5 models offer dramatically lower pricing than Anthropic's Claude alternatives, with Claude Opus 4 costing up to 50 times more for output than GPT-5's most affordable tier.
Microsoft has introduced a $30 Copilot add-on for Office and OpenAI launched a $200 ChatGPT Pro plan, with higher price points aiming to ensure revenue covers hefty compute costs. However, premium pricing strategies risk customer pushback in competitive markets.
GitHub introduced "premium requests" in April 2025, imposing rate limits when users switch to AI models beyond the base model, with Copilot Pro users receiving 300 monthly premium requests and additional requests costing $0.04 each.
"This paywall is necessary to prevent runaway costs from a small set of very heavy users," Kulikova explained. "But the paradox is that when they do put in usage controls, users don't like the experience."
Dependency Risks
The challenges extend beyond foundational model providers to companies building atop them. Anthropic's revenue concentration shows approximately $1.2 billion of the company's $4 billion revenue milestone came from just two customers: coding applications Cursor and GitHub Copilot, representing nearly a quarter of income.
Cursor sends 100% of its revenue to Anthropic, which uses that money to build Claude Code, a competitor to Cursor, with Cursor deeply unprofitable even before Anthropic added service tiers that increased enterprise pricing.
Kulikova described this dynamic as symptomatic of industry-wide problems. "If a GPT wrapper startup puts a low price on AI usage or offers a beefy premium and doesn't limit expensive features, a minority of users can generate costs that will scale exponentially," she said.
Rare Success Cases
Despite widespread struggles, certain applications demonstrate sustainable economics. Companies with control over which models to use and workflow depth can achieve better margins, as classic SaaS was essentially a wrapper over databases and public cloud yet sustained 70-80% gross margins.
Drawing on her LegalTech experience, Kulikova identified successful AI businesses as those working with large volumes of text-based data in industries including accounting, HR, sales, and legal services. "The vast majority of GenAI startups that work are apps that work with large amounts of text-based data and documents in various industries," she said, citing examples of tools that stitch together contract data with invoices and automate communications between parties.
Anthropic's Claude Code has quickly generated over $500 million in run-rate revenue with usage growing more than 10x in three months, demonstrating strong demand for developer-focused AI tools.
However, Kulikova cautioned that even successful applications face limitations. "It's not going to make billions in revenue, but it is a sustainable and viable business model," she said of contract automation tools.
The Product Management Litmus Test
Based on her experience scaling products from concept to launch, Kulikova proposed a framework for evaluating AI business viability: "A traditional SaaS startup with AI features for tasks that can be automated—working off of traditional SaaS benchmarks, not bubbled up AI metrics."
The critical question, she argued, is whether products deliver value independent of AI capabilities. "The point where you know that your product is valuable is when it is valuable without AI and when it could solve the real problem without AI," Kulikova explained. "If there is an AI component in it and it makes something better or faster, fantastic. But AI should not be the defining factor of the product."
Market Outlook
Industry executives remain divided on AI's economic trajectory. Microsoft CEO Satya Nadella suggested AI could boost global GDP growth to 10% annually, while Wharton School researchers project a more modest 1.5% increase in productivity and GDP by 2035.
A McKinsey survey showed that of companies reporting cost reductions from AI, most had savings of less than 10%, while companies with revenue increases saw gains of less than 5%.
Technology executives warned against an agentic AI hype cycle, cautioning that investors shouldn't expect profitability in the next three to five years.
AI Supernova startups reach approximately $40 million ARR in their first year and $125 million in their second year, but often with fragile retention and thin margins, contrasting with "Shooting Star" companies that grow from $3 million to $100 million over four years with strong product-market fit and healthy margins.
Strategic Imperatives
Industry experts recommend AI companies maintain gross margins above 50% and an LTV:CAC ratio of 3:1 or higher, warning that ratios below 1:1 mean "selling dollars for 90 cents".
The key question for viable AI products is whether use cases require top models for every request or can meet quality bars by routing most traffic to cheaper models and bursting to frontier models when needed.
Kulikova's conclusion reflects her years navigating enterprise software economics: "It's easy to get a user to try a product with big promises and even more so with FOMO and scare tactics. What's really difficult is to deliver value—value that a few can replicate."
She pointed to traditional metrics as the ultimate arbiter of success. "Retention and conversion rates—those two things will outlive blown up VC rounds or hype waves," Kulikova said. "The hype won't last forever, but boring, profitable businesses will."
As the sector matures, companies face mounting pressure to demonstrate sustainable unit economics rather than relying indefinitely on venture capital. The coming years will likely separate viable businesses from those dependent on continued investor enthusiasm and media hype—a distinction product managers like Kulikova, with experience building platforms that generate hundreds of millions in recurring revenue, understand intimately.
Sources
No comments:
Post a Comment