Aaron Levie has a knack for making enterprise software sound exciting. The co-founder and CEO of Box, the cloud content management company he started from a college dorm room, has spent the past two years positioning his firm at the center of the AI transformation sweeping corporate America. But his latest claim isn’t about what AI can do. It’s about what AI will cost.
Almost nothing, he says. And soon.
In a recent interview with Business Insider, Levie predicted that by 2026, the cost of using AI tokens — the fundamental units that large language models process when generating text, analyzing documents, or answering questions — will drop to near zero. His argument is straightforward: the economics of computing have always trended toward deflation, and AI inference costs are following the same trajectory, only faster. “The cost of intelligence is going to be essentially free,” Levie told the publication.
That’s a bold statement from a CEO whose company is actively building AI-powered products that depend on those very tokens. But Levie isn’t alone in noticing the trend. And the implications, if he’s right, would ripple through every corner of the technology industry — from the hyperscalers selling compute to the startups burning through venture capital to train and deploy models.
To understand why Levie’s prediction carries weight, consider the math. OpenAI’s GPT-4, when it launched in March 2023, charged developers $0.03 per 1,000 input tokens and $0.06 per 1,000 output tokens. By early 2025, newer models from OpenAI, Anthropic, Google, and a constellation of open-source competitors have driven those prices down by orders of magnitude. Google’s Gemini Flash models, for instance, offer pricing that would have seemed absurd just eighteen months ago. The pattern mirrors what happened with cloud storage, bandwidth, and processing power over the past two decades — except the cost curve is steeper.
Levie’s company is betting heavily on this trajectory. Box has integrated AI capabilities across its platform, allowing enterprise customers to use large language models to summarize contracts, extract metadata from documents, and automate workflows that previously required human review. The cheaper the underlying inference becomes, the more aggressively Box can embed AI into every feature without worrying about margin erosion. It’s a virtuous cycle for a SaaS company: falling input costs mean higher gross margins on AI-powered features, which means more investment in AI-powered features.
But here’s where it gets complicated.
Not everyone in Silicon Valley shares Levie’s optimism about the cost trajectory, at least not without significant caveats. While the per-token price of running a query against an existing model has indeed plummeted, the total cost of AI deployment for enterprises remains substantial. There’s the expense of fine-tuning models on proprietary data. There’s the infrastructure required to handle retrieval-augmented generation, or RAG, which lets AI systems pull from a company’s own documents rather than relying solely on pre-trained knowledge. There’s the cost of guardrails, monitoring, and compliance — none of which get cheaper just because tokens do.
Sam Altman himself has acknowledged this tension. OpenAI’s strategy depends on making its models widely accessible at low cost while simultaneously spending tens of billions of dollars on the data centers and custom chips needed to serve them. The company reportedly lost money on its API business through much of 2024, subsidizing usage to build market share. Whether that’s sustainable depends on whether efficiency gains in hardware and model architecture can outpace the explosive growth in demand.
Nvidia’s Jensen Huang has offered a different framing. He’s argued that as AI gets cheaper per unit of compute, total spending on AI will actually increase — because cheaper intelligence means companies will find more uses for it. Economists call this Jevons’ paradox: when a resource becomes more efficient to use, consumption of that resource often goes up, not down. Applied to AI tokens, it suggests that even if the price per token approaches zero, the aggregate spending on AI infrastructure could continue climbing for years.
Levie seems to understand this dynamic intuitively. In the Business Insider interview, he didn’t argue that companies would spend less on AI. He argued that the unit economics would shift so dramatically that AI would become a default feature of every software product rather than a premium add-on. For Box, that means AI isn’t a separate SKU to be sold at a markup. It’s table stakes. The company that embeds intelligence most deeply into its product wins.
This is a strategic bet with real consequences. Box reported fiscal fourth-quarter results in March that showed revenue growth of around 5%, a pace that some analysts consider modest for a company trying to ride the AI wave. The stock has had a bumpy ride over the past year, reflecting uncertainty about whether Box can translate its AI ambitions into accelerating growth. Levie’s argument about collapsing token costs is, in part, a pitch to investors: trust us, the economics are about to get dramatically better.
Recent developments in the broader AI market lend some credibility to his timeline. In February 2025, Chinese AI lab DeepSeek released models that matched or exceeded the performance of leading Western competitors at a fraction of the training cost, sending shockwaves through the industry. The implication was clear — the moat around frontier AI models is narrower than many assumed, and competition will continue driving prices down. DeepSeek’s emergence rattled Nvidia’s stock and forced a reassessment of how much companies really need to spend on AI infrastructure to remain competitive.
And then there’s the open-source factor. Meta’s Llama models, Mistral’s offerings, and a growing roster of community-driven projects have created a floor under how much any proprietary model can charge. If a company can run a capable open-source model on its own infrastructure for pennies per query, the ceiling on commercial API pricing drops accordingly. This dynamic is accelerating. Every month brings new open-weight models that close the gap with proprietary leaders.
For enterprise software companies like Box, Salesforce, and Microsoft, the collapse in token costs creates both opportunity and risk. The opportunity is obvious: embed AI everywhere, make products stickier, charge for outcomes rather than compute. The risk is subtler. If intelligence becomes a commodity — as cheap and abundant as electricity — then the competitive advantage shifts entirely to data, distribution, and workflow integration. Companies that own the data layer, as Box does for document management, may be well positioned. Companies that were banking on AI itself as a differentiator may find their moat evaporating as fast as their token costs.
Microsoft offers an instructive comparison. The company has poured more than $13 billion into OpenAI and built Copilot features across its entire Office and Azure product lines. But Microsoft isn’t pricing Copilot based on the marginal cost of AI inference. It’s charging $30 per user per month for Microsoft 365 Copilot — a price that reflects the value of the productivity gains, not the cost of the underlying tokens. If token costs fall to near zero, Microsoft’s AI margins could become extraordinary. Or competitors could undercut the pricing, forcing a race to the bottom.
Levie’s prediction also raises questions about the venture-backed AI startup market. Hundreds of companies have raised billions of dollars on the premise that they can build AI-powered products with defensible margins. If the core technology — the ability to process and generate language, analyze images, write code — becomes essentially free, then what exactly are these startups selling? The answer, presumably, is domain expertise, proprietary data, and user experience. But those have always been hard to monetize at venture-scale returns, and the bar gets higher when the underlying technology is commoditized.
There’s a historical parallel that’s worth examining. In the early 2000s, the cost of hosting a website collapsed. Companies like Amazon Web Services, launched in 2006, made compute and storage available at commodity prices. That didn’t kill the software industry. It supercharged it. An entire generation of SaaS companies — Salesforce, Workday, ServiceNow, and yes, Box — was built on the assumption that infrastructure costs would keep falling. The winners weren’t the companies that provided the cheapest compute. They were the companies that built the best products on top of cheap compute.
Levie is essentially arguing that we’re at a similar inflection point with AI. The infrastructure layer is commoditizing. The value will accrue to companies that use nearly free intelligence to solve specific, high-value problems for customers. For Box, that means being the company that helps enterprises manage, search, and act on their unstructured data — contracts, invoices, medical records, legal filings — using AI that costs almost nothing to run.
Whether he’s right about the timeline is another question entirely. “Near zero by 2026” is aggressive. It assumes continued exponential improvement in model efficiency, sustained price competition among providers, and no major supply constraints on the specialized chips that power AI inference. Any disruption in the semiconductor supply chain, any consolidation among model providers, any regulatory intervention could slow the trajectory. The AI chip market remains dominated by Nvidia, and while AMD, Intel, and a host of startups are pushing alternatives, a true commodity market for AI compute hasn’t materialized yet.
Still, the direction of travel seems clear. Token costs are falling. They’re falling fast. And the companies that plan for a world of nearly free AI — rather than one where intelligence remains expensive and scarce — will likely be better positioned than those that don’t.
Levie, for his part, seems energized by the prospect. He’s been one of the most vocal enterprise tech CEOs on social media, posting frequently about AI developments and Box’s product roadmap. His enthusiasm isn’t new — he’s been a perpetually optimistic pitchman since founding Box in 2005 at the age of 20. But there’s a specificity to his current predictions that suggests he’s seeing something in the data that gives him confidence.
The question for Box’s investors, customers, and competitors isn’t really whether token costs will approach zero. The question is what happens to the enterprise software market when they do. Because when the cost of intelligence drops to nothing, the value of the problems you solve with that intelligence becomes everything.
WebProNews is an iEntry Publication