Last month, Chinese AI company DeepSeek unveiled two developments through its new model R1, quietly redefining the economics of artificial intelligence. It delivered top-tier performance at just 1/40th the cost of previous models, not long after the company’s V3 large language model, released in December, remarkably slashed training costs by more than 90%.
First, DeepSeek revealed that asking AI models to narrate their reasoning—an approach that researchers call chain-of-thought prompting—improved both accuracy and efficiency. Second, DeepSeek used AI to generate its own data sets, employing zero humans to label data. While there’s belief DeepSeek’s accomplishments did not come as cheaply as claimed, the advancements are still ushering in a new era of AI economics.
AI’s cost structure is changing. And changing fast. Skyrocketing performance per dollar has massive implications for startups, enterprise adoption, and infrastructure investments. Shifting fundamentals may upend market forces in favor of scrappy startups, who finally have the opportunity to outpace incumbents while boosting margins.
Tech giants, for their part, have collectively poured over $100 billion into infrastructure development (and counting), and have to determine how to generate returns on their whopping investments and sustain an algorithmic advantage over leaner market entrants. Given rapid rates of change, the message to tech giants and startups is clear and uniform: Move quickly to harness advancements while they last, or risk being left behind.
AI before and after DeepSeek
Before DeepSeek, the landscape looked like this: Startups struggled to compete with tech giants on infrastructure spending, which demands enormous resources. AI incumbents have spent tens of billions of dollars per quarter on capital expenditure to build large data centers, benefiting disproportionately from AI advances. The combination of access to huge volumes of data, a large concentration of PhDs on staff, and algorithmic advances only possible through skill thrust goliaths to the forefront of the industry. Long-standing distribution networks further allowed these giants to funnel their products to existing customers, rapidly advancing their technologies through feedback loops.
But David’s sling has now become as powerful as Goliath’s army of PhDs. Training costs have dropped by 95%—by a factor of 20—in 2025 alone, diminishing infrastructural advantages. Over the last three years, inference costs have reduced almost a thousand-fold; accordingly, we should expect an additional hundred-to-thousand-fold reduction to come. Algorithmic advantages now last about 45 to 100 days and should only plummet further.
When training costs are no longer the key constraint, inference workloads—how efficiently AI models run in real-time applications—become the new focal point. We’re entering a phase where smaller, cheaper models can offer comparable capabilities to their larger predecessors. They can also be run on less powerful GPUs, extending the lives of older GPUs that can be repurposed. If smarter AI products can be delivered at a fraction of the cost, startups finally have the opportunity to outpace incumbents, while boosting margins.
Efficient staffing redoubles challenger advantages. With the need to hire an entire cohort of PhDs to stand up a competitive AI outfit greatly reduced—if not fully erased—startups can develop, refine, and distribute models at a fraction of the human-capital cost incurred by incumbents. Further, by largely existing at the application layer, challengers benefit from increased margins, the same way they leveraged unit-economic improvements to cloud computing 15 years ago.
The shifting winds aren’t only a boon for startups. They should frighten companies like Nvidia, whose stock dipped 12% after DeepSeek’s announcement, although it has since rebounded. Chip developers face heightened risks as demand shifts from training-focused hardware to more efficient inference solutions. The rise of consumer-grade neural processing units (NPUs) could further accelerate this shift, enabling AI models to run locally on devices like smartphones and laptops.
AI spending
What’s bullish for challengers is bearish for AI’s big-tech overlords. In almost knee-jerk fashion, AI giants pointed to the national-security implications of DeepSeek’s dominance in an effort to buoy support for their developmental techniques, glossing over the fact that domestic researchers, including at Stanford, have already replicated or outmatched DeepSeek’s standards. Looking to the future, enterprises that have poured GDPs’ worth of capex into data-infrastructure projects may justifiably ask: Have all the hundreds of billions of dollars that have gone into AI models been a waste of money? If the cheap thing works as well as the expensive thing, why pay so much?
Historical trends suggest most AI advances have indeed come as a result of scale and excessive spending on that scale. Transformer architecture worked brilliantly because of overtraining: training more than what was thought to be algorithmically ideal. Newer examples show that we are able to attain the same level of performance at significantly less cost, but even with DeepSeek-esque efficiencies, hyper-scalers’ push to innovate on new models will require ever-larger data centers, and host ever-ballooning inference costs.
Tech giants are not standing still. We’re already seeing an arms race to replicate and surpass DeepSeek’s achievements with Google’s Gemini models, Microsoft’s Azure AI Foundry, and Meta’s open-source LLaMA vying for dominance. Open-source models will likely play a pivotal role. Meta CEO Mark Zuckerberg has highlighted the importance of personalized AI—models tailored to individual users’ needs, cultures, and preferences. This vision aligns with broader trends in AI development: smaller, specialized models that deliver high performance without the need for massive cloud infrastructure.
AI startups gain leverage
Meanwhile the bifurcating goals of open-source and closed-source goliaths provide additional leverage for challengers. Open-source models created by Meta and others will continuously look to compete and reduce the overall cost within the ecosystem. Closed-source models, meanwhile, will try to extract a premium for better technology. Startups can play these two factions’ competing priorities off each other to generate the best price performance on a per-use basis, all the while improving their margins.
The message to every outfit of every size is clear: Move quickly to harness particular advancements at their disposal—market dynamics, compute, talent—or fail. Progress timelines are getting shorter and shorter. Where it once took years or months to set a new performance benchmark, DeepSeek’s advances suggest it now takes just 41 days. Innovation is intensifying at a rapid rate, and the margin for error is narrowing rapidly, too.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
Read more:
- The 8 things most people missed about DeepSeek
- IBM CEO: DeepSeek proved us right—AI is not about big, proprietary systems
- AI ‘prompt and pray’ hasn’t cut it in the enterprise, but we’ve found the missing puzzle piece. Mass deployment is next
- The most underreported and important story in AI right now is that pure scaling has failed to produce AGI
