The case so far has been built from theory and history. Cheap intelligence erodes barriers to entry. When a fundamental input becomes cheap, competition intensifies, margins compress, and the businesses that relied on that input’s scarcity are exposed.
The telegraph did it to information brokers. Containerization did it to protected manufacturers. The internet did it to anyone whose margins depended on controlling distribution.
Now the tense shifts from past to present. The argument is no longer that this could happen, or that history suggests it will. It is measurable, right now, across multiple categories of evidence.
Twenty-four hours
On February 2, 2025, OpenAI released Deep Research, a feature that uses an AI agent to conduct extended multi-step research across the web and synthesize its findings into a structured report. It scored 67% on the GAIA benchmark, a test of real-world research capability. The previous open-source state of the art was 46%.
Twenty-four hours later, Hugging Face published an open-source clone. A team had set itself a one-day deadline, built a working system, and scored 55% on the same benchmark.
That was not the only response. Jina AI’s CEO began work within hours of OpenAI’s announcement and had a functional version merged into code by the following morning. Within the same week, LangChain, Together AI, and Firecrawl all published their own implementations. Five open-source alternatives to a product that had existed for less than seven days.
This was not an isolated case. On January 20, 2025, DeepSeek released R1, a reasoning model that matched OpenAI’s o1 at a fraction of the cost. Eight days later, Hugging Face launched Open-R1, an effort to reproduce the entire training pipeline.
Google’s NotebookLM went viral in September 2024 with its audio overview feature, surging from 652,000 monthly visits to over 3 million in a single month. A data scientist at Singapore’s GovTech built an open-source equivalent in a single afternoon. Meta released its own version, NotebookLlama, about six weeks later.
Anthropic launched Computer Use on October 22, 2024, the first public API for AI-controlled computer interaction. Ninety-three days later, OpenAI shipped Operator. Four days after that, Browserbase had an open-source alternative on GitHub.
The pattern accelerates even within the span of a few months. NotebookLM was cloned in weeks. DeepSeek R1 was cloned in days. Deep Research was cloned in hours.
The wrapper economy
The frontier-model cloning cycle is the most visible expression of this dynamic. The application layer is the most pervasive.
The industry calls them “wrappers”: products built on top of foundation model APIs, adding interface, workflow, or domain-specific logic. Market Clarity, a research firm tracking the space, estimates roughly 35,000 AI wrapper applications globally, with ten to fifteen new ones launching every day. Only two to three thousand have meaningful traction. Ninety percent shut down within eighteen months. Between 3 and 5% ever surpass $10,000 in monthly recurring revenue.
These numbers do not describe a quality problem among startups. They describe an environment where the cost of building a functional competitor has collapsed so far that everyone tries, and the resulting competitive density makes sustained differentiation nearly impossible. The wrappers are not interesting as individual businesses. They are interesting as a signal: when the barrier to creating a product drops below the threshold of a weekend project, the number of attempts explodes and the survival rate converges toward zero.
Jasper and the Pony Express
The trajectory of Jasper AI is the sharpest illustration.
In October 2022, Jasper raised $125 million at a $1.5 billion valuation. It was a content generation tool built on OpenAI’s API, and at the time it looked like a category winner. Annual revenue had roughly doubled year-over-year to $80 million. The company projected $250 million in ARR by the end of 2024.
Then ChatGPT launched on November 30, 2022.
By summer 2023, Jasper had revised its revenue forecasts down by at least 30%. On September 28, both the CEO and CTO stepped down. The internal valuation was cut 20%. Revenue, which had peaked around $120 million in 2023, fell to an estimated $55 million in 2024, a decline of more than 50%. The company that had been valued at $1.5 billion less than two years earlier was now struggling to justify a fraction of that figure.
Jasper’s collapse was not caused by a better product. It was caused by a cheaper, more general one. ChatGPT did not specifically target Jasper’s market. It made the entire category of AI writing tools contestable by anyone with a browser. What Jasper had sold for $80 a month, OpenAI gave away for free, and the competitive advantage that had justified a billion-dollar valuation evaporated in months.
The Pony Express was the clearest case of what happens to a business built on a form of scarcity that suddenly disappears. The Pony Express operated for nineteen months. It grossed $90,000 and lost $200,000. It was never viable once the telegraph crossed the continent.
Jasper is the Pony Express of the AI era. It was not a bad business. It was a business built on a temporary advantage that disappeared the moment the underlying capability became broadly available. The difference is speed: the Pony Express had nineteen months. Jasper’s window from peak valuation to leadership departure was eleven.
Time-to-clone as leading indicator
The examples above are not anomalies. They are data points on a curve, and the curve is compressing.
When the time between a product’s launch and the emergence of functional competitors drops to hours or days, first-mover advantage has a half-life measured in weeks. The four frontier-model labs shipped their latest models within twenty-five days of each other in late 2025. Google’s Gemini 3 reportedly triggered an emergency at OpenAI that accelerated GPT-5.2’s release by weeks. At the application layer, functional clones now appear faster than most companies can execute their marketing plans.
Time-to-clone is the leading indicator of competitive intensity. It measures how quickly the market can mount a credible response to anything new. When that interval compresses toward zero, the traditional model of building a product, establishing a market position, and extracting rent over time stops working. You are not competing for market share. You are competing against the speed at which your margins can be arbitraged away.
The second category of evidence is price.
Time-to-clone measures how fast competitors appear. Price measures how fast the value of appearing first erodes. And the pricing data, from the foundation model layer down through every market that depends on cognitive work, tells a single story: costs are falling at rates that compress competitive advantage before it can be monetized.
The stack is in free fall
In February 2025, Sam Altman published a blog post titled “Three Observations.” The second observation: “The cost to use a given level of AI falls about 10x every 12 months.” He compared it to Moore’s law, which changed the world at 2x every 18 months. “This is unbelievably stronger.”
The data supports the claim. Epoch AI, in a March 2025 study tracking inference costs across multiple benchmarks, found prices falling at a median rate of 50x per year. Since January 2024, the median has been closer to 200x per year, with a range across tasks from 9x to 900x. The cost of matching GPT-4-level benchmark performance dropped from roughly $20 per million tokens at GPT-4’s launch to $0.40 by mid-2025. A 50-fold decline in two and a half years.
Those are aggregate trends. The competitive dynamics within the AI industry itself are sharper.
When DeepSeek released R1 in January 2025, it priced reasoning capabilities at $0.55 per million input tokens. OpenAI’s comparable model, o1, charged $15. A 27-fold difference for roughly equivalent performance. Nvidia lost $589 billion in market capitalization in a single day.
What followed was a pricing rout. OpenAI priced its next reasoning model, o3, at $2 per million tokens, an 87% cut from o1. GPT-5 launched at $1.25. Anthropic cut Claude Opus by 67%, from $15 to $5. OpenAI shipped GPT-4.1 with a 26% price reduction it explicitly framed as a response to competitive pressure.
By February 2026, the pricing hierarchy had inverted. DeepSeek V3, a general-purpose model, was available at $0.14 per million input tokens. GPT-5, OpenAI’s flagship, cost $1.25. Gemini Flash-Lite, Google’s budget tier, was $0.075. Chinese AI labs had entered a pricing war of their own, with multiple companies launching new models around China’s Spring Festival, each undercutting the last. DeepSeek offered 75% off-peak discounts.
OpenAI, the company that defined the commercial AI market, faces projected losses of $14 billion in 2026 and is seeking $100 billion in additional funding. It is simultaneously the most important company in AI and the clearest illustration of what happens when the product you sell is being commoditized faster than you can monetize it.
This is not a temporary price adjustment but the pattern of a commodity market forming in real time. Stanford’s 2025 AI Index captured the compression: the performance gap between the top model and the tenth-best narrowed from 11.9% to 5.4% in a single year. Sixteen labs produced models exceeding GPT-4’s capability within two years of its release. The Menlo Ventures enterprise survey showed OpenAI’s market share halving from 50% to 25% in eighteen months. The intelligence layer is heading where compute always heads: toward a utility priced at marginal cost.
The cascade downstream
None of this would matter much if the pricing collapse stayed in the AI stack. It does not.
In customer support, human agents cost $5 to $15 per inquiry. AI chatbots handle the same interactions for $0.50 to $2. McKinsey reported a 50% reduction in cost per call globally for organizations deploying AI agents. Vodafone documented a 70% cost reduction. Klarna’s AI chatbot, in its first month of operation in early 2024, handled two-thirds of all customer service conversations, work previously done by 700 full-time agents. Resolution time dropped from 11 minutes to under 2. Repeat inquiries fell 25%.
The Klarna case is instructive for what happened next. Customer satisfaction dropped 22%. The CEO admitted the company “went too far,” that cost had been “a predominant evaluation factor” over quality. By spring 2025, Klarna was rehiring human agents and building a hybrid model. The pricing pressure was real. The human cost of over-automating was also real. Both facts can coexist. The relevant point for competitive dynamics is not that AI replaced every agent. It is that the cost floor for acceptable customer service moved permanently downward. No company competing against Klarna can now staff a 700-person call center and pass the cost to customers.
In translation, the gap is wider. Human translators charge $0.08 to $0.40 per word. AI translation runs at $0.001 to $0.05. Enterprise localization workflows using AI with human review are reporting 60% cost savings with 97% accuracy. In content creation, the average AI-generated marketing post costs $131 versus $611 for human-written, a 4.7x difference.
In software development, the pressure is structural. Developer freelance rates dropped from $75 per hour in 2022 to $48 per hour by mid-2025, a 36% decline. Ramp, a corporate expense management platform, published data showing that the share of total company spending going to freelance labor marketplaces fell from 0.66% in late 2021 to 0.14% by mid-2025. Over the same period, spending on AI model providers rose from zero to nearly 3%. More than half the businesses that had used freelancers in 2022 had stopped entirely. The substitution ratio: for every dollar companies stopped spending on freelance developers, they spent three cents on AI. Fiverr’s stock fell more than 60% in a year. The company cut 25% of its workforce. Its market capitalization dropped below its 2019 IPO valuation.
Legal services are a lagging indicator, but the gap between cost savings and client pricing is becoming visible. Law firm AI adoption jumped from 19% in 2023 to 79% in 2024. Lawyers report saving the equivalent of 32 working days per year. Yet only 6% of firms have passed savings to clients through reduced rates. Thirty-four percent are charging premium rates for AI-enhanced work. Bloomberg Law’s assessment: “AI Does Little to Reduce Law Firm Billable Hours.”
That asymmetry will not hold. Ninety-four percent of in-house legal teams are actively exploring alternative service models. When the buyer side of a market knows the seller’s costs have dropped, pricing pressure follows. The legal profession is watching what happened to customer service, translation, and freelance development. It is not yet experiencing the same compression. It will.
The pattern in real time
Each of the historical precedents followed the same pricing arc. The telegraph compressed commodity price spreads by more than a third within months. Containerization cut cargo handling costs by 97%. The internet drove stock trading commissions from $49 to zero.
The AI pricing collapse has the same structure but operates on a compressed timeline. Foundation model prices are falling 50x per year. Downstream services are seeing 60 to 97% cost reductions. The businesses that depended on the scarcity of cognitive work are watching their pricing power erode quarter by quarter.
And unlike previous cycles, where the pricing collapse moved through industries sequentially over decades, this one is hitting customer support, translation, content, software development, and professional services simultaneously. The input that got cheap is the one that every knowledge-work industry depends on.
The third category of evidence is not about speed or price. It is about what a competitive company looks like when both collapse at once.
Speed tells you how fast competitors appear. Price tells you how fast margins erode. The third category of evidence is structural: the minimum viable competitive entity has changed.
The numbers that don’t make sense
In November 2025, Cursor, an AI code editor, hit $1 billion in annual recurring revenue. It had taken twenty-four months to get from $1 million to $1 billion. The company had roughly 300 employees.
That trajectory would have been impossible five years ago. Not unlikely. Impossible. A billion dollars in recurring revenue from 300 people breaks every scaling model the software industry has ever produced.
Cursor is not an outlier.
Lovable, an AI application builder, went from $1 million to $100 million in annual recurring revenue in eight months with about 45 employees. Bolt.new hit $20 million ARR in sixty days with roughly 15 people. Gamma, a presentation tool, reached $100 million ARR with approximately 50 employees while remaining profitable for fifteen consecutive months. Midjourney generates an estimated $500 million in annual revenue with about 160 employees and zero external funding.
Put these side by side and the pattern is hard to miss. Jeremiah Owyang calculated average revenue of $3.48 million per employee across a sample of AI-native companies, compared with $611,000 for leading SaaS companies. A 5.7x gap.
One person, eighty million dollars
The most striking data point comes from the smallest possible team.
Maor Shlomo built Base44 alone, using AI-assisted development. The platform reached $3.5 million in ARR. Six months after launch, Wix acquired it for $80 million in cash.
One person. Six months. Eighty million dollars.
This is not a feel-good founder story. It is evidence of what happens to market structure when one person can build what previously required fifty.
Shlomo is not a one-off. Carta’s data shows solo founders rising from 22% of new startups in 2015 to 36% in the first half of 2025. The reason is mechanical, not inspirational: when AI can generate code, design interfaces, write documentation, and handle customer support, the fixed costs of company formation collapse. What used to require headcount now requires API subscriptions.
Y Combinator’s Summer 2025 batch made the pattern visible at scale. Over 90% of the batch was AI-focused, the most homogeneous cohort in the accelerator’s twenty-year history. More than a dozen startups building AI code editors alone went through YC between 2022 and 2024. The accelerator that built its reputation on funding diverse ideas was now funding variations on a single thesis.
What this means for competition
Any one of these companies is a good story. Together, they are a structural argument about who gets to compete.
If it takes 300 people to build a billion-dollar-revenue product, then every market that previously required thousands of employees to compete in is now contestable by teams an order of magnitude smaller. If a solo founder can reach acquisition-level revenue in six months, then the number of potential competitors in any given market has increased by orders of magnitude.
Baumol’s contestable markets framework predicted exactly this. When the cost of mounting a competitive challenge falls, incumbents lose pricing power even if no challenger actually enters. The mere possibility of entry disciplines the market.
But the challengers are not staying on the sidelines. They are entering.
The funding data confirms it. AI startup investment hit $202 to $211 billion in 2025, roughly half of all global venture capital. By Q3 2025, AI deals accounted for 63% of every VC dollar invested. CB Insights counted more than 170 AI agent startups in March 2025 alone, and that was one subcategory. The capital is not concentrating in a few firms. It is dispersing across thousands of small teams, each capable of building products that three years ago would have required hundreds of engineers.
The incumbents notice
The enterprise software industry can read these numbers too.
The high-growth share of enterprise software companies declined from 57% in 2023 to 39% in 2024. AlixPartners projects it will fall to 27% in 2025. Public SaaS median net dollar retention dropped from 123% in Q3 2022 to 108% by Q2 2025. These are not recession metrics. Revenue is still growing at most of these companies. But the growth rate is decelerating, and the reason is not macroeconomic. New entrants are fragmenting the market faster than incumbents can expand within it.
The simplest way to see it is in the multiples. Median SaaS revenue multiples fell from above 7x at the start of 2025 to below 5x by year’s end. The public markets are not pricing these companies for the growth they used to deliver. They are pricing them for a world where 45-person teams can build competitive alternatives in months.
This is the structural consequence of the speed and pricing evidence. When competitors appear in hours and prices fall 50x per year, defensibility erodes on both sides at once. And when a 45-person company can reach $100 million in ARR, the incumbent with 10,000 employees is not ten times more capable. It is ten times more expensive.
The fourth and final category of evidence is what happens when all of this occurs across every market at once.
Speed measures how fast competitors appear. Price measures how fast margins erode. Structure measures how small a team can be and still compete. The fourth category is what happens when all three operate across every market at once.
The marketplace as census
A useful proxy for competitive density is the number of products fighting for the same users.
The VS Code marketplace now lists more than a thousand AI-powered extensions. Over 90% were published in the last two years. G2, a software review platform, catalogs 544 AI writing assistant products, a category that grew 170% in listings between 2022 and 2023. CB Insights mapped more than 400 AI agent startups across 16 categories by early 2026, more than doubling since March 2025. Enterprise AI agent and copilot revenue reached an estimated $13 billion in 2025, up from $5 billion the year before.
These are not fringe categories. They are the core workflows of knowledge work: writing, coding, customer interaction, data analysis, sales, legal research. Every one of them is seeing competitive entry from hundreds of new products, most built in the past eighteen months, many by teams smaller than a single department at the incumbents they compete against.
The capital flows match. The OECD, using a broader definition of AI firms than most private trackers, pegged AI startup funding at $258.7 billion in 2025, representing 61% of all global venture capital. That share had been 30% three years earlier. At the top end, 58% of 2025 AI funding went into megarounds of $500 million or more. Fifteen companies raised above $2 billion. At the other end, thousands of small teams raised seed rounds to build products that compete directly with incumbents orders of magnitude larger.
This is not a gold rush. Gold rushes produce a few winners and many losers in a single territory. This is competitive entry across every territory at once, with the cost of staking a claim falling toward zero.
The SaaSpocalypse
The public markets made these dynamics concrete in February 2026.
On January 30, Anthropic published eleven open-source plugins for Claude Cowork to GitHub. The plugins enabled an AI agent to handle contract review, compliance checks, sales preparation, legal intake, and internal research, capabilities that had previously required multiple SaaS subscriptions costing $500 to $900 per month combined, now bundled into a $100 monthly AI plan.
The market reaction arrived on February 3. Thomson Reuters fell 16% in a single session, its largest intraday drop in company history. Salesforce and ServiceNow each dropped 7%. LegalZoom fell 20%. By mid-February, a five-day selloff that Wedbush labeled “Software-mageddon” had erased more than $800 billion from the S&P 500 Software Index. By month’s end, total enterprise software market capitalization had declined by more than $1 trillion.
The damage was broad. ServiceNow fell 51% from its all-time high. Atlassian dropped 57% year-to-date. Intuit fell roughly 40%. The iShares Expanded Tech-Software ETF dropped from $117 to $82, a 30% decline. The Morgan Stanley SaaS basket traded at 18x forward earnings, its cheapest level on record, against a historical average above 55x over the prior decade. Jefferies traders christened it the “SaaSpocalypse.”
Palantir’s CTO, on the company’s Q4 earnings call the same week, claimed its AI tools could reduce complex ERP migrations from years of work to as little as two weeks. The implication was blunt: if agents can write and manage enterprise software, the companies selling it on a per-seat basis are exposed.
Bank of America called the selloff overblown, comparing it to the DeepSeek panic of January 2025. But the underlying arithmetic was real. AlixPartners warned that AI agents could trigger a $500 billion collapse in enterprise software revenue, on the logic that if one agent does the work of ten human users, per-seat licensing math stops working. The pricing model that sustained SaaS for two decades, subscriptions scaled to employee headcount, faces a structural problem when the relevant headcount is no longer human.
The pattern completes
The four categories of evidence form a single picture.
Competition arrives in hours or days. Prices fall 50x annually at the infrastructure layer and cascade through every downstream market. A 45-person company can reach $100 million in recurring revenue. And every market that depends on cognitive work is experiencing these pressures at the same time.
This is the pattern from the telegraph, from containerization, from the internet, playing out in present tense. The input became cheap. The barriers fell. The competitors entered. The margins compressed. The difference: information, physical goods, and distribution each affected industries that depended on that specific input. Intelligence on tap affects every industry that depends on cognitive work. Which is nearly all of them.
The evidence here is not a prediction. Time-to-clone is measured in hours. Pricing declines are measured in orders of magnitude per year. Company formation is measured in weeks. Market crowding is measured in hundreds of competitors per category. And the $1 trillion in enterprise software value erased in February 2026 is a market repricing these measurements in real time.
The question is no longer whether hypercompetition is arriving. It is what the economy looks like when competition operates at the speed and intensity of high-frequency trading, the most compressed competitive environment the financial world has ever produced.