Bank AI budgets exceed 100 billion dollars, but can't match coding's explosive growth
Banks and pharmaceutical companies have raised AI budgets to their highest-ever priority, but Zhang Lu believes enterprises must traverse two stages—automation and optimization—to truly transform workflows, making year-one explosive growth like coding impossible.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Embedded in workflows, Claude has become enterprise infrastructure
Zhang Lu says the prospectus figure that struck her most wasn't revenue growth but customer composition: roughly 11,000 to 12,000 enterprise customers spending over $100,000 annually, with about 2,000 spending over $1 million yearly. This reveals Claude isn't merely a trial tool but deeply embedded in workflows—functioning like infrastructure. The result: far faster upsell velocity and opportunity.
— Zhang LuEnterprises cut closed-source adoption from 57 to 14 percent in six months
A report from two weeks prior showed large enterprises' adoption of cutting-edge, closed-source models collapsed from 57% to 14% in half a year, shifting toward open-source deployment. AT&T publicly committed 60-70% of internal AI budgets to open-source; Nvidia's Jensen Huang also publicly backs the open-source ecosystem. But Zhang Lu doesn't expect this to squeeze Anthropic and OpenAI from the market—AI is permeating the entire economy, expanding the pie itself. They'll simply have more competitors dividing it.
— Zhang LuFTX collapse forced liquidation of SBF's Anthropic stake for debt
Anthropic's early financing was rocky. In May 2023, Spark Capital made a non-consensus bet at Anthropic's ~$4.5 billion valuation when markets weren't sure whether they could support a second frontier model company. More dramatically: FTX founder SBF had invested in Anthropic too, but after FTX's bankruptcy, his shares were largely auctioned off for debt repayment. Zhang Lu noted that had FTX not collapsed, those investment returns could have covered much of what SBF needed to pay back.
— Zhang LuUS bank's 2026 AI budget exceeds 100 billion dollars annually
Zhang Lu disclosed that a major US bank's Chief AI Officer told her the bank's AI budget this year has already reached over 100 billion dollars—shocking even her. She observes a pattern: once one company in an industry commits heavily to AI, the rest race to catch up fast. Companies fear falling behind AI-savvy competitors more than being disrupted by startups.
— Zhang LuEnterprise AI transformation requires automation first, then optimization—years not quarters
Zhang Lu divides enterprise AI into two stages: Business Automation (embedding AI in existing workflows for cost cuts and efficiency gains), then Business Optimization (lifting revenues, showing in financial statements). Most enterprises don't want to scrap systems and rebuild; they want agents to keep working with their current software. She estimates stage one alone taking 3-5 years to roll out broadly, with stage two taking longer. Don't expect AI to transform the entire economy in one or two years.
— Zhang LuA 500 million monthly bill taught enterprises to ration AI use
Zhang Lu relayed that a consulting firm heard an unnamed large enterprise's single-month Claude bill suddenly spike to nearly 500 million dollars, shocking their CFO. This unsustainable burn quickly triggered cost controls: major tech companies shifted from early encouragement to burn tokens to now strictly rationing usage. This partly explains why enterprises favor open-source models—closed-source pricing is prohibitive.
— Zhang LuPharma's real bottleneck is raising clinical success rates, not discovering targets
Regarding Anthropic and OpenAI's bet on AI for Science (drug discovery, mathematical proofs), Zhang Lu thinks near-term revenue contribution is unlikely, though reasoning ability will improve. She pushes back on AI pharma narratives: even if AI expands target discovery from 2-3 targets to hundreds, those don't directly become drugs—clinical trials consume the cash. She's bullish on using AI to identify, from hundreds of targets, the few with highest clinical success rates, lifting success above 50-60%—that's genuinely valuable to the industry.
— Zhang LuPhysical AI failures mean deaths, not hallucinations—regulation must come first
Zhang Lu recently spent considerable time thinking about AI regulation—her UN AI for Good board just held a 2+ hour session on it. Language model hallucinations were once funny; Physical AI failures risk human lives and catastrophic accidents. CTOs she meets are anxious, insisting that regulation and technical safeguards must be embedded before scaling physical AI—a prerequisite no high-regulation industry like healthcare can bypass.
— Zhang LuIn their own words · checked verbatim
He already has, to numbers I recall, roughly eleven thousand-plus, possibly up to twelve thousand enterprise customers spending over a hundred thousand dollars annually.
他已经拥有了超过我记得数字 大概是一万一千多家 可能到一万两千家年消费超过十万美元以上的客户
Zhang Lu4:02
Large enterprises' adoption of cutting-edge AI models has already dropped from 57 percent to 14 percent.
现在大的企业 对于这个B元模型的 或者说前沿AI模型的 使用率 已经从57%降到了14%
Zhang Lu10:03
Once one company in an industry moves, the rest of them catch up very fast.
那只要一个行业有一家动起来 剩下几家企业追上速度就很快
Zhang Lu34:15
I heard from their chief AI officer that the bank's AI budget this year has already reached over a hundred billion dollars—having such a massive AI budget as a bank is truly surprising.
我从他们的首席AI官听到了 他在AI的今年的预算 是已经达到了100多亿美金 他是一个银行 他有这么大的一个AI的预算 这个是让人非常惊讶的一件事情
Zhang Lu35:15
Previously you could only discover two or three targets, and now you can discover hundreds—but those hundreds still can't become drugs; you still have to do clinical trials.
以前可能你只能发现 两三个target 那现在能发现 几百个target 但是你这几个target 也不能形成药啊 你还是要去做临床 做clinical trial
Zhang Lu56:27
With physical AI, if anything goes wrong, it could harm human lives and cause massive accidents.
physical AI的这个情况 余境下的话 如果AI出现任何问题 那可能就会人的生命受到伤害 甚至说出现非常巨大的一个事故
Zhang Lu1:08:33
Figures
| Enterprise customers spending over $1 million annually | approximately 2,000 | 4:02 |
| Net Revenue Retention rate | 900% | 18:08 |
| API (usage-based) revenue YoY growth, H1 2026 | 1200% | 30:15 |
| Subscription revenue YoY growth, H1 2026 | 1690% | 31:15 |
| Enterprise adoption of cutting-edge/closed-source models | from 57% to 14% | 10:03 |
| AT&T's internal AI budget allocation to open-source | 60-70% | 11:04 |
| Non-core Anthropic employees' options eligible for vesting on IPO day | 10% | 13:06 |
| Anthropic's valuation when Spark Capital invested | approximately $4.5 billion | 19:09 |
| Single-month Claude bill for major enterprise | nearly $500 million | 38:15 |
Glossary
- Physical AI
- AI systems requiring operation of physical equipment and interaction with the physical world.
- Net Revenue Retention (NRR)
- A metric measuring year-over-year spending growth from existing customers.
- Business Automation / Business Optimization
- Two sequential stages of enterprise AI deployment—first reducing costs and improving efficiency, then lifting revenue.
- AI for Science
- Using AI to assist scientific research including mathematical proofs and drug-target discovery.
How to listen
Investors sizing AI market growth and tracking B2B startup opportunities, and practitioners wanting to understand enterprise AI budgets and actual deployment speeds.
The opening discussion of lock-up design and valuation ranges—lower information density for non-primary-market practitioners.