State of Startups and AI 2025

What will definitely happen by the end of 2026: AI agents shipping code directly into your production environment, voice replacing text for most business communication, inference falling below a cent per million tokens, or a WALL-E-like future of leisure? Sarah Guo opens with this audience poll, specifying real production environments rather than playgrounds. After a brief negative response, she lands on the first option: engineers expect agents to ship.

Sarah Guo helped start Conviction roughly two and a half to three years before this talk, just before ChatGPT. She came from Greylock; her partner Mike Vernal came from Sequoia. Their reason for starting another venture firm was that AI changes not only technical capabilities but also the businesses, product decisions, and operating problems built around them. Their investments span applications such as Cursor, Cognition, Harvey, and OpenEvidence, model companies such as Mistral and Thinking Machines, and infrastructure such as Baseten.

After about a decade investing before Conviction, Guo sees user uptake unlike anything in her earlier experience. AI engineering and product development have proved harder than people hoped, yet useful products can grow exceptionally quickly. Guo reports seeing companies reach $10 million, $50 million, and $100 million in run rate unusually quickly from zero, without specifying a common elapsed time. The practical question is whether a product will keep working for its customers over the long term. Actual usage is a better guide to that question than arguments about an AI winter, endless summer, or daily market movements.

The opportunity Guo calls Cursor for X starts with capability: what can models now do that a useful application can package? Reasoning provides another route to scaling intelligence with compute. Its application value appears in high-stakes decisions where showing work matters, sequential problems that require several connected steps, and problems that benefit from systematic search. These resemble ordinary knowledge-work tasks, not merely benchmark exercises.

Reasoning also feeds enthusiasm for agents, which Guo punctuates with a Steve Ballmer-style repetition of the word. Her operational definition is more useful than the slogan: an agent plans steps, incorporates AI, takes ownership of a task, and retains a goal in memory. It can try hypotheses, backtrack, and use tools such as search or other models. The implementation may be simple or sophisticated; the distinguishing feature is a system that acts on a task, behaving more like a colleague than a conversational endpoint.

Conviction’s Embed grant program provides a window into what founders are attempting, including user and revenue information submitted with applications. Guo says Embed receives thousands of applications annually and that agent startups reached 50% of applications over the preceding year. She also sees real deployments working, not just proposals using agent terminology.

Voice, video, and image generation are becoming businesses as well as demonstrations. Guo describes HeyGen, Eleven, and Midjourney as passing $50 million in ARR. The HeyGen clip she plays makes the capability change tangible: successive presenters deliver a joke about self-expression and being banned from three coffee shops, a reflection on hands that hurt or heal, a business update about Q3 projections, and a product recommendation. These are lines spoken inside the demonstration, rather than Guo’s own business claims.

The advance is not simply a face reading text. HeyGen can depict clones of real people or invented people whose gestures and expressions reflect the content and emotion of their speech. Multiple models work together to produce the result. Guo jokes that the final presenter has influenced her enough to buy the bunny, even though she does not know what it is: an illustration of how expressive delivery changes the experience.

A common objection is that enterprises do not currently possess much voice, video, or image data. But the value of collecting data changes when software can structure and understand it. Making these modalities useful creates a reason to capture more of them. Video already carries a large share of everyday communication because it conveys so much at once.

Guo expects voice to enter business workflows first because the communication pattern already exists. Medical consultations and lead generation already involve people speaking; the constraint has been scaling those interactions. Other modalities become more attractive as control improves and costs fall. Her forecast is continued improvement throughout the model layer, rather than an imminent halt imposed by a data wall.

Improving capabilities arrive in an increasingly competitive market. Guo invokes Sam Altman’s characterization of the previous year’s model as a commodity: a striking prospect when that model is already useful. Guo’s pricing illustration puts GPT-4 at $30 and then $2 per million tokens over about 18 months, with unnamed distilled versions around $0.10. These figures express her historical cost argument, not a like-for-like price schedule: the talk does not distinguish input from output tokens or identify the distilled models. The historical Structured Outputs announcement prices the named successor GPT-4o, gpt-4o-2024-08-06, at $2.50 per million input tokens and $10 per million output tokens.

The OpenRouter chart shows competition in actual model selection: Google is green and Anthropic yellow. Guo describes Claude gaining share against OpenAI and Google returning strongly with Gemini. The chart represents OpenRouter traffic, whose users are already inclined toward multiple models; it excludes substantial usage sent directly to OpenAI. It therefore illustrates diversity within that routing market rather than total industry market share.

Competition also includes prospective entrants SSI and Thinking Machines, whose researchers pursue different technical approaches. Guo points to DeepSeek base and reasoning models as reasonably competitive at a claimed fraction of training cost, without supplying a matched cost-and-performance comparison. For an application builder, the direction is useful: open-source and proprietary suppliers compete for the application’s business. Tools such as OpenRouter and inference platforms such as Baseten support a product strategy that expects more than one model.

Coding applications show that substantial value can accumulate above the model layer. Guo reports Cursor growing from $1 million to $100 million ARR in 12 months, reaching half a million developers, and initially having no salespeople. She describes Cognition, which began with more autonomy, as already the top committer at many companies.

At the time of the talk, Guo describes a proposed $3 billion OpenAI acquisition of Windsurf, interpreting the prospect as evidence that labs value application businesses. That statement should not be read as a completed OpenAI transaction; the later Cognition announcement identifies a different acquirer. Guo also reports Lovable and Bolt each reaching $30 million ARR in a handful of weeks by helping non-engineers build software.

Code had several advantages as an early application domain:

• Structured language: Code is logical text with explicit structure. Much engineering involves sophisticated boilerplate; producing an API endpoint or React component need not require general intelligence.
• Deterministic validation: Compilation, tests, and execution provide automatic checks of proposed work.
• Research investment: Researchers view coding as important to AGI, making it a priority for benchmarks, training, and data collection.
• Workflow familiarity: Engineers built tools for engineers, with an intimate understanding of how their customers work.

The last advantage is the most transferable. Builders in another industry need both AI knowledge and a precise understanding of the customer’s problem, so they can redesign the workflow around model capabilities.

Cursor for X does not mean attaching one model to a new profession. In Guo’s description, Cursor uses different models for diffs, merges, and file embeddings. It manipulates and packages context, prompts models carefully, and lets users standardize repetitive work with Cursor rules. She also describes retrieval improving through better coverage and freshness as the product is used.

That machinery sits inside a familiar interface. VS Code users retain recognizable interactions and shortcuts. Green additions and red deletions let them inspect a proposed change before accepting it, while responsiveness keeps review from becoming frustrating. A small proposed change illustrates why this representation matters:

 export function greeting(name: string): string {
-  return `Hello, ${name}`;
+  return `Hello, ${name.trim()}`;
 }

The reader can see exactly which behavior would change and which code remains intact. The proposal is not an applied edit; the interface gives the user a concrete decision to make. This is part of the product’s value, alongside the model that generated it.

Guo jokes that this is a very substantial wrapper, invoking a roughly $14–15 billion valuation. Her burrito analogy makes the allocation explicit: imagine 80% wrap and 20% filling, with an open market supplying interchangeable fillings. If the model is increasingly substitutable, the surrounding company can still create value through context, integration, and the experience of getting work done.

The corresponding design procedure is concrete:

1. Start with known customer needs. Do not make users explain things the product should already know about their industry.
2. Collect and package context automatically. Include information from sources beyond the user’s natural-language request.
3. Choose models for particular steps. Orchestration means using the right model at the right point in the workflow.
4. Present results in a useful interface. Outputs need thoughtful presentation, not merely a text box.

Guo discourages another generic chat interface except as a learning exercise. She sees continuing work for graphical interfaces, product taste, and careful implementation.

“The prompt is a bug, not a feature.” The provocation is about user effort: a product that understands its workflow should not repeatedly require users to reconstruct its context. The best interactions can feel like mind reading because the application has already gathered what it needs. Software companies packaged workflows in 2015, and they still do in 2025; AI changes what those workflows can accomplish.

Beyond coding, Guo sees a surprising pattern: some conservative, comparatively low-tech industries appear to be adopting AI especially quickly. Conviction calls this the AI leapfrog effect. The opportunity is to solve an existing professional problem well enough that adoption need not wait for the customer to become an AI expert.

Guo reports Sierra resolving 70% of customer-service queries for its customers, naming SiriusXM and ADT as examples. The talk does not define the measurement period, customer cohort, or resolution criteria. Guo reports Harvey exceeding $70 million ARR two years in and describes its AI as becoming essential to legal competitiveness.

OpenEvidence places intelligent medical-research search at the point of clinical decision-making. A clinician supplies a medical identification number to access the product. Guo reports OpenEvidence reaching one-third of U.S. doctors weekly, with the average user using it daily. Those frequency claims are her account; the contemporaneous investor evidence supports the broad one-third adoption scale rather than establishing both frequencies.

The shared mechanism is customer understanding applied to a real problem. Guo also points out that Bret at Sierra chairs OpenAI’s board and that OpenAI was Harvey’s seed investor. Their involvement undercuts the idea that people close to the labs necessarily see every application company as a dispensable thin wrapper.

Returning to Embed, Guo adds that the agentic share of applications doubled over the preceding year. But enthusiasm for full automation does not make copilots obsolete. Her observations of what drives revenue leave her thinking that augmentation remains underrated.

The Iron Man suit is a useful comparison because autonomy is a spectrum. It augments Tony Stark while also being able to fly on command and perform some tasks without him.

Guo’s practical concern is that tolerance for hallucinations, failure, and unreliability falls sharply as latency increases. Waiting longer makes a disappointing result more frustrating. For many domains, the path is to build augmentation that people adopt, then extend its autonomy as capabilities improve: first get someone wearing the suit.

Guo points to Conviction’s requests for startups, then sketches a range from obvious model fit to work that was previously impractical. Legal text generation fits the capabilities of language models directly. A different category appears when machines can actively gather information from people—what Mike Vernal describes as machines interrogating humans.

Consider talking to every customer rather than only the top 5% by contract value, or proactively investigating the root cause of every alert instead of firefighting. These are proposed workflow changes, not reported automation outcomes. The design question is what an organization would do differently if it had an army of compliant, infinitely patient knowledge workers. Abundant labor can change the coverage of a process, rather than merely speed up its existing steps.

Some valuable problems require more than reorganizing information already available online. In robotics, biology, materials science, and physics simulation, the answer—and the reasoning needed to obtain it—may not be in Common Crawl. Progress requires clever data collection and often interaction with physical systems: atoms as well as bits.

Guo tentatively connects the reasoning that succeeds on math Olympiad problems with the ability to navigate molecular space. Her broader interest is the overlap between machine learning and other scientific disciplines: shared architectures can become useful when paired with experts and data from a different field. These opportunities may be unfamiliar to software builders, but they reach fundamental scientific and societal questions.

What makes an application defensible as the models beneath it improve? The warning is familiar: stay out of the labs’ path rather than picking up pennies in front of a steamroller. Guo’s answer is execution. The ability to keep delivering a better product is available to application builders, not reserved for model developers.

Cursor did not invent code completion, the underlying model, or even its product surface. Guo credits its ability to ship a better experience faster than competitors could copy it, earning developer loyalty at least for the current period. That qualification matters: execution produces a lead that must be maintained.

Jasper provides her contrasting example. Guo cites its first-mover advantage, brand, and $125 million raised, but characterizes the initial product as prompts in a text box supported by strong SEO. She says ChatGPT quickly undermined that first iteration. The comparison concerns a vulnerable product approach, rather than a claim that the entire company disappeared.

A substantial workflow product can accumulate customer trust, adoption, and context that competitors do not automatically possess. Much of the data needed to improve these experiences is not readily available today. Building the experience that earns access to it is an opportunity for application companies. Staying ahead therefore means continuing to improve the workflow and the customer relationship, not merely securing an early place in a category.

Guo describes AI as moving from dial-up toward broadband. Her mobile-platform analogy emphasizes that transformative applications did not all arrive with the platform itself. She places Instagram, Uber, and DoorDash four, five, and six years after the iPhone, respectively—her illustrative chronology rather than a precise company-founding timeline. The relevant distinction is between noticing a platform change and reimagining an experience around it.

AI also keeps changing the conditions for building. Guo likens it to receiving a meaningfully different iPhone every year: a new model, a capability breakthrough, or a cost reduction to one-tenth can reopen an opportunity. A product idea that was previously too expensive or unreliable may become practical, while an established product may need to change again.

Engineers received these capabilities early, but their current usage does not define the whole market. Guo cites roughly 40% coding usage from Anthropic’s Economic Index. That should not be read as 40% of all AI use: the March 2025 report studies Claude.ai conversations, and its computer-and-mathematical occupational category is broader than coding; the exact report behind her rounded figure is unspecified. More fundamentally, even a large coding share would not mean coding accounts for the same share of economic opportunity. The work ahead is translation: understanding another profession well enough to turn model capabilities into an experience its practitioners can use.