How AI Companies Can Become Anti‑Fragile in the Token Economy

01

At Nvidia's GTC 2026, Jensen Huang described data centers not as file warehouses but as token factories, stating that an AI company's throughput and token‑generation speed will directly translate into next‑year revenue. This framing positions tokens as the new electricity and oil of the AI era, giving them strategic and rigid demand.

The token economy reshapes AI business models: selling APIs, memberships, or ads is portrayed as less promising than selling tokens. Capital markets have responded with a wave of listings and record‑size financing for AI firms that adopt the token model.

However, the enthusiasm masks a growing risk: token prices are increasingly being treated as a simple “must rise” narrative. As intelligent agents replace chatbots, compute costs rise exponentially, prompting many AI companies to raise token prices. Examples include OpenAI and Anthropic’s multiple price hikes, Chinese platforms such as ByteDance, Alibaba, and Tencent adjusting plans, and DeepSeek’s 30% increase for its GLM Coding Plan.

The article argues that a “sell because demand is high” logic is unsustainable. Competitive markets tend toward supply‑demand balance; early movers may enjoy temporary pricing power, but as the market matures, raising prices becomes harder and can create friction between AI providers and enterprise customers.

02

Drawing on Nassim Taleb’s *Antifragile*, the author stresses that AI companies must develop resilience to unpredictable “black‑swans” rather than relying on fragile, price‑driven revenue streams.

DeepSeek illustrates a technology‑driven cost‑reduction approach. Its V4 model delivers output prices that are 99% lower than GPT, Claude, or Gemini, and its token‑consumption per token is only 27% of the previous V3.2 version, with KV‑cache usage reduced to 10%. After a brief price increase, DeepSeek launched a free‑trial plan offering 1,500 calls per five‑hour window.

SenseNova 6.7 Flash‑Lite from SenseTime adopts a multimodal architecture that eliminates the visual‑to‑text intermediate layer, cutting token consumption by up to 60% in information‑search scenarios. The company also offers a free‑trial token plan with a generous call quota.

Both firms pursue a “technology‑driven cost reduction” (技术降本) strategy, lowering token price and consumption to make AI services affordable and sustainable, rather than relying on price hikes.

03

The shift toward productivity‑oriented AI is presented as the key to anti‑fragility. DeepSeek focuses on “AI + productivity,” optimizing its models for coding assistants and agent workflows. SenseTime builds a comprehensive office‑skill suite, enabling enterprises to assemble custom agents for tasks such as data auditing, financial analysis, and dynamic pricing recommendations.

Capital markets have begun to value this productivity focus. OpenAI’s valuation is reported at $852 billion, Anthropic at $900 billion, while DeepSeek’s recent financing exceeds 500 billion RMB with a post‑money valuation of 3.5 trillion RMB. SenseTime’s full‑stack strategy—combining compute infrastructure, large‑model R&D, and AI applications—positions it as a long‑term contender despite a current valuation gap.

In conclusion, the article asserts that AI companies can achieve anti‑fragility by lowering token costs through technical innovation, centering product development on tangible productivity gains, and building diversified, resilient revenue streams that can withstand market volatility.