More molecules, same failures: the crack in AI drug design
China's AI-for-drug-discovery crowd spent Sunday at the Pujiang Innovation Forum arguing about the thing benchmarks don't measure — what happens after the model proposes. And the country's biggest banks are quietly building a lending product around the same question: is AI usage real enough to lend against?
"The more you generate, the more you fail." That was Ding Sheng, founding dean of Tsinghua's school of pharmaceutical sciences, at the forum's Future Medicine track on September 13, and the sentence captures the awkward phase generative chemistry has entered: the front end of the funnel got enormous, and the wet lab behind it did not get cheaper. Chen Kaixian, a CAS academician from the Shanghai Institute of Materia Medica, put the upside in national terms — China approved 11 new-target, new-mechanism drugs in the first half of 2026, all of them domestic discoveries, against 11 for all of 2025 of which only four were homegrown. The room's consensus was that the number flatters the technology: participants complained that AI drug design still runs short of clean data, explainable models and affordable experimental validation, and that the field only changes character once the loop from computation to automated experiment and back into the model actually closes. A new AI-pharmaceutical innovation consortium was announced at the forum to try to do exactly that. Our read — this is the same defect every agentic system shows up close: generation is cheap, verification is the product. Pharma just said it out loud with a patient-safety price tag attached, and the honest metric nobody is reporting yet is success rate, not molecules per week.
A month after Guangdong issued the first "token loan," the product has a catalogue — and a fraud debate. Bank of China's Guangzhou branch now sells three variants: lending against compute-supply contracts for firms that train and host models, against consumption for firms that burn tokens on their own products, and against settlement flow for API, data-labeling and fine-tuning vendors, capped at 30 million yuan per borrower over three years. Construction Bank's Guangdong arm took a different cut of the same chain, approving a 30 million yuan line for a Nansha smart-hardware firm on contracts and receivables, while ICBC's Shenzhen Longgang branch priced a 5 million yuan line for an embodied-AI company whose spray-painting robots get smarter one token batch at a time. The collateral is consumption itself — monthly token counts, compute contracts, API logs; one Guangzhou short-drama studio told reporters that a finished minute of ordinary AI video burns about 1,500 yuan of tokens and a high-quality minute about 4,000, and used its usage history to win a 5 million yuan three-year line. Haizhu alone has committed 28 million yuan across early borrowers in AI video, e-commerce marketing and tool development. Regulators set the frame in July, when the central bank and eight other ministries told lenders to move credit assessment away from fixed assets and toward data. The catch, admitted by the banks themselves: no mature risk model yet, and repeat-selling idle tokens, doctored call logs and padded consumption are already known market tricks — so every application still gets cross-checked against contracts, cash flow, credit records and the founder's history. Our take — the substance behind the branding is that inference bills became auditable before AI revenue did, which is why lenders trust a log over a business plan. We covered the first loan in August — Bank of China's first 'token loan' ties credit to AI usage.
What to watch: whether any Chinese AI-pharma consortium publishes a validation-throughput number, and whether a token-loan default shows up before the product scales past Guangdong.
Is usage telemetry a legitimate basis for a loan, or just a metric waiting to be gamed? Tell us in the comments.
