Anxiety doesn't come out of thin air. The most famous vertical applications, such as Cursor from Anysphere for programmers, Harvey for lawyers or OpenEvidence for doctors, they grew rapidly because they were able to wrap general AI models into specific, profitable processes.
Their valuations speak for themselves. Anysphere announced a $900 million round in June. at a valuation of USD 9.9 billion, and according to later reports, investors even considered entering at approximately USD 30 billion. OpenEvidence raised $200 million in October. at a valuation of USD 6 billion. In turn, Harvey, which in May was negotiating financing at USD 5 billion, was recently valued at USD 8 billion. This is proof that the market believes in the advantage of specialized AI over the promise of a single, all-powerful AGI.
Meanwhile, laboratories are not waiting. OpenAI introduced an agent that “thinks and acts” independentlyhandling complex tasks on your computer and on the network. “The Guardian” described it as a personal assistant capable of booking, shopping and working on files. Additionally, OpenAI has replaced the Assistants API with a new Responses API designed for agents, emphasizing its “chat to execution” ambition. In practice, this is entering the territory of many automation and productivity improvement start-ups.
Check also: Excel turbocharged with AI. Tracelight transforms the work of analysts thanks to artificial intelligence
AI models are looking at startups
Anthropic doesn't get left behind. After introducing the Claude 3.x line, the company developed Claude Code, which is full-fledged coding tools, as well as Artifacts, which allows you to create and share small applications without code. If a business's advantage was previously based on “AI for programming”, it is now competing with the AI model provider itself.
The economic layer is also tightening. In a high-profile piece, columnist Ed Zitron described Anthropic's costs on the AWS cloud, suggesting that the scale of cloud bills prompted the company to change the prices of its serviceswhich hit big clients like Cursor. The Economist, meanwhile, noted that even though labs retort that varying service levels are normal practice, the impression remains that vendor pricing and policies could make it significantly more difficult for many startups and AI application developers to thrive at any time.
Paradoxically, there is also a race to the bottom in terms of technology. Price pressure is growing, and increasingly efficient open source models and distillation tools mean that good enough systems can be run cheaper. The market is in control offensive of cheap models from China, and distillation helps cut inference costs (distillation is training a smaller model to imitate a large one, making it work almost as well but faster and cheaper to use). For applications, this means that model providers also become their alternatives.
Read also: She threatened and blackmailed employees. Anthropic's newest AI is incredibly powerful
Attempt to protect margin
It's no wonder that start-ups protect margins by using model routing, i.e. automatically directing each query to the model that best (or cheapest) fits a given task, so that optimize quality and cost. They also have mixed technology solutions: simple queries go to cheaper, open-source models, and expensive cases go to the most powerful APIs.
An ecosystem of intermediaries like OpenRouter allows you to switch providers virtually on the fly. This is not only cost control, but also negotiation leverage with the largest laboratories.
Start-ups may feel threatened by AI models, but this does not mean inevitable failure. The creators of companies such as Sierra talk about the real battlefield hidden under the water – not in the general chat layer, but in complex, non-portable business processes. Bret Taylor emphasizes that in AI, there is a transition from subscriptions to pay-for-result, and the advantage comes in the often most boring, but also the most expensive and consistent activities of a given industry. This is where operational data, integrations, compliance and quality assurance come into play and cannot be cloned by one generic AI model.
The financial market also does not assume the omnipotence of laboratories. According to HSBC analysis, by 2030 LLM providers can only get about 30 percent. from worth USD 1.3 trillion. IT services pie supported by AI. The rest will go to companies that will “package” the models into specific products and processes. If this projection holds true, applications—not models—will be where most of the value is created.
See also: OpenAI completes the transformation into a commercial company. Now the earning will start
Start-ups need to be careful
From the perspective of a start-up founder, the conclusions are sobering. Dependence on one API is too big a risk today — not only in terms of pricing, but also in terms of product, since the same supplier may tomorrow launch a function that competes with our core module or restrict access in specific jurisdictions, as when introducing agents outside the EU.
Therefore, it is worth designing your architecture with the possibility of quick model replacement, and consciously building things that are difficult to copy. That is, your own data corpus, result validation mechanisms, audit trails, user training and, finally, reputation for great results in a specific field. There lies a moat that even the best overall architecture cannot overcome.
There are also reasons for optimism. The success of Cursor, Harvey and OpenEvidence shows that specialization and obsession with quality in specific professions can generate hypergrowth even in the shadow of giants. This isn't just a matter of an interface over someone else's API. It is the art of transforming generative AI into a reliable work infrastructure with guaranteed accuracy, legal liability, industry certification and insurable risk. Generic models may provide the “brain”, but the nerves and muscles of the organization are built by the application. At least for now.
Author: Grzegorz Kubera, journalist of Business Insider Polska
