刘一鸣
Key Views & Dialogues
E225 | Silicon-Based Employees Are Here, Wiping Hundreds of Billions from SaaS Market Cap: How Will AI Reshape Organizational Structures?
Enterprise software is being repriced from seats to auditable outcomes, with Bairong cutting contract review from 56 minutes to 4 and positioning RaaS as a market at least dozens of times larger than SaaS. Private data, complex workflows and dedicated models may buy SaaS time, but Agents are already absorbing recruiting, review and unstructured-content work, leaving humans with judgment, sign-off and accountability.
View summary (8)
- The core of this software-stock re-rating is that enterprises are shifting what they pay for from seats to auditable business outcomes. The episode says global software stocks lost nearly $1T in a week after Anthropic launched 11 enterprise plugins on Claude Cowork; 张韶峰 argues that the technology did not suddenly leap forward, and that OpenClaw and these plugins were merely “the last straw.” Licenses will not disappear immediately, but “the curtain has definitively risen on a new paradigm”; the old model’s decline and eventual extinction are only a matter of time. He also recalls that in 2021 investors asked 百融 to adopt American-style SaaS, while he argued that China could skip SaaS and go straight to TaaS (Transaction); investors later suggested BaaS (Business, business-level services).
- Private data and complex workflows are the main factors that can extend SaaS’s lifespan, while industry know-how confined to human minds is not a durable moat. General-purpose models struggle to replicate applications without access to private-domain knowledge; a workflow may contain 20 steps, and 张韶峰 uses repeated multiplication by 0.99 or 0.98 over 25 steps to show how error probabilities accumulate, potentially falling below an enterprise’s tolerance threshold and giving traditional vendors perhaps 3 years of “redemption.” 刘一鸣 immediately argues that it may not even be 3 years, asking whether that estimate is too optimistic; 张韶峰 does not directly confirm this, instead turning to a third category of barriers. He still believes that almost no Chinese SaaS companies meet these conditions, and that only a small fraction of American companies might.
- RaaS expands the software budget to cover salaries, business outsourcing and end-to-end outcomes, and 张韶峰 estimates that its market is at least dozens of times larger than SaaS. 百融 cut internal contract review from 56 minutes of human labor per contract to 4 minutes, with a silicon-based employee handling the remaining 52 minutes; pricing can therefore be based on roles, hours or units, with costs potentially just one-third to one-fifth those of carbon-based employees. AI BPO can go further and price on end-to-end outcomes; 张韶峰 compares this with traditional sales channels taking 20%, 30% or 40% of sales, without saying that 百融 itself uses those fixed rates.
- For traditional software giants, their customers and installed-base revenue are both a moat and the heaviest obstacle to self-disruption. Moving from one-time On-premise revenue plus 15%–20% maintenance fees to SaaS, where upfront revenue is lower and renewals uncertain, creates an inherent conflict of interest; AI transformation also faces legacy customers with much lower tolerance for errors. Startups, by contrast, are “barefoot”: they have no legacy revenue to cannibalize, and early users are more willing to accept mistakes. By the time the new paradigm is broadly accepted, the “new king” may already have a mature product and organization.
- AI Agents are being treated as governable, measurable and retireable workers rather than merely software features. 张韶峰 says 百融’s more than 1,000 carbon-based employees correspond to 200K–300K silicon-based employees, putting the silicon-to-carbon ratio at roughly 130–150:1; silicon-based employees have organizational structures, employee IDs, email addresses, tenure, JDs, KPIs and designated trainers. When performance is poor, they may first improve themselves, then be retrained by carbon-based partners or “retired from service”; when they perform well, “all a silicon-based employee needs is electricity and chips,” while ultimate responsibility and economic incentives remain with humans. He also once built a silicon-based assistant to handle internal approvals, and the need was later standardized into a company-wide silicon-based employee.
- Automation does not necessarily mean layoffs in equal numbers, but whether workers can migrate into new roles will determine how the gains are distributed. 百融 originally used 50 carbon-based employees to serve 2,500 small-business customers, each contributing RMB100K–500K in annual revenue; it later shifted to 18 silicon-based employees and 5 carbon-based employees. The remaining 45 people were not laid off: they learned to produce agents, moving from a cost center to a profit center, and revenue increased. 张韶峰 retained an important caveat: “I’m afraid not every enterprise can do this.”
- The value stack for enterprise Agents may sit across dedicated models, AI PaaS, IAVs and Agent Stores rather than being captured by a single general-purpose model vendor. 百融 says its dedicated models handle more than 95% of problems, with the remaining roughly 2%–5% of corner cases routed to general-purpose models; AI PaaS fills in the common capabilities traditional ISVs need to become Independent Agent Vendors. If an ecosystem partner creates ¥100 of revenue, it could notionally take ¥70 and the platform ¥30; the principle is, “If you don’t make money, we don’t charge you either.”
- Recruiting, legal, business and tax services show that Agents will first absorb search, communication, unstructured-content production and multi-step workflows, while humans retain review, judgment and accountability. Recruiting Agents can understand the capabilities behind “microservices” rather than merely search the keyword, and candidates may be required to use Cursor, Claude Code, Codex or Trae on-site; 百鉴 cut internal contract review from 56 minutes to 4 minutes and generated cross-border consulting reports in tens of minutes. 张韶峰 expects that in 3 years, professionals may “be left with only one respectable job: review and sign-off,” but that signature will still represent credentials, judgment and final responsibility.
Vol.71 A Quick Chat About GPT-5 and U.S. Power—with 刘一鸣/101 Weekly
GPT-5 suggests reinforcement learning, synthetic data, and pretraining reuse can extend scaling laws, while inference-token demand may grow exponentially despite falling unit costs. OpenAI still needs To B revenue beyond To C engineering to fund expanding CapEx, while data-center spending could revalue the U.S. power chain, though the critical point remains unverified.
View summary (7)
- GPT-5 did not prove that the data wall has disappeared; it showed that reinforcement learning can extend the scaling law through new routes. 庄明浩 believes Ilya was right about the original-data bottleneck he identified in September 2024; what has changed is the combination of synthetic data, pretraining reuse, and post-training—a self-lifting loop in which “the left foot steps on the right and flies upward.” Mathematics, coding, and medicine are precisely the fields where reinforcement learning can deliver clear gains.
- Compute demand will continue shifting from training to inference, and falling unit costs will not reduce total demand. 庄明浩’s order-of-magnitude estimate is that token costs fall by roughly an order of magnitude every 12 months, while token consumption and required compute also rise exponentially over the same period—“it evens out.” Inference tokens and compute are not simply linearly related; time and per-use costs also matter, so the problem cannot be solved with a single equation.
- GPT-5’s launch fell short of expectations, but product engineering is still filling gaps on the To C front. Reducing hallucinations, improving routing, removing sycophancy, and implementing finer-grained engineering changes are all improving the product experience; better tool use, longer tasks, and higher success rates point to a more reliable agent. 庄明浩’s view is that “the main battlefield is still To C.”
- OpenAI has to push harder into To B because consumer subscriptions are unlikely to cover expanding CapEx on their own. ChatGPT’s weekly active users rose from roughly 300M–400M at the start of the year to 700M, but even with further growth, weekly actives multiplied by conversion and then by monthly ARPU of $20 or $200 still produces an estimable number; enterprise implementation and government contracts offer larger marginal gains, while competition from Anthropic and cloud providers must also be confronted directly.
- The broad scaling law still holds, but capability growth has shifted from “feeding in more data” to combining pretraining, post-training, and reinforcement learning. 庄明浩 believes this R&D path was still working at least through Q3 2024; GPT-5’s gains in tool use, task length, and success rates fit the L3 agent phase. As for the next milestone and how L4 will be achieved, “we outsiders may simply have no way of knowing.”
- The re-rating of U.S. power assets is being driven by data-center CapEx reaching a critical point for the power system, not by GPT-5’s single release. Several leading companies are now spending roughly $100B per quarter; 庄明浩 relayed a Morgan Stanley report saying annual investment could reach $3T by 2029 or thereabouts, leading him to speculate that “the entire U.S. power ecosystem could be rewritten.”
- The opportunity in power expansion extends beyond generation equipment to the full chain of production, transmission, distribution, and auxiliary infrastructure. 刘一鸣 is watching GEV’s steam gas turbines, steam turbines, SMRs, hydropower, and related businesses, noting that its order book runs through 2029; both speakers drew a clear line around their judgment—they are extrapolating from AI demand and CapEx, and do not yet know whether the data has reached a critical point or exactly what needs to change.
Vol. 69: Revisiting the Stargate Project — with 刘一鸣/101 Weekly
The bearish compute trade reversed after DeepSeek, with NVIDIA’s market cap surpassing $4T, but Stargate’s bottleneck has shifted from whether to add compute to who can fund infrastructure bills worth tens of billions annually. Oracle monetizes data-center construction and operation more directly, while SoftBank is betting on its future, ARM and Japan’s AI deployment; OpenAI targets $10B–$13B against Oracle’s $30B annual contract, leaving financing and cost coverage as key risks.
View summary (7)
From NVDA’s 17% one-day plunge after DeepSeek’s shock to NVIDIA’s market cap breaking $4T and the Nasdaq and S&P 500 hitting new highs, the market seemed to prove in the first half of 2025 that concerns over compute spending could be set aside for now. Agents, reinforcement learning, multimodality, and AI coding continue to consume compute, while the bearish narrative epitomized by Microsoft’s decision to cut two data centers has faded; 庄明浩’s summation was, “It seems everything is back.” But as consensus again moves to an extreme, new risks may be brewing.
Stargate was not knocked off its compute-driven logic by DeepSeek; the real bottleneck has shifted from “Should we keep piling on compute?” to “Who pays for it?” OpenAI is expected to generate only $10B–$13B in revenue for the year, yet signed a $30B-a-year data-center services agreement with Oracle; whether Microsoft, Google, Amazon, Facebook, Oracle, or another source of capital ultimately picks up the tab is more unsettled than the technology trend itself. “People are no longer discussing the trend… someone will always foot the bill.”
The projects now being placed under the Stargate umbrella no longer appear to be the clearly defined standalone entity announced in January. The original plan called for outside investors and shareholders to provide funding, Oracle to build the infrastructure, and a formal operator to participate; SoftBank is not part of that original plan, while OpenAI and Microsoft’s interests and financing arrangements, SoftBank’s financing of OpenAI, SoftBank’s own fundraising, and its Stargate commitments overlap and nest inside one another, forming “a knot that cannot be untangled.”
SoftBank is not an ordinary financial investor; it is simultaneously betting the group’s future, ARM, AI deployment in Japan, and its relationship with the Japanese government. 庄明浩 did not call failure “fatal,” saying only that it would be “quite troublesome”; Masayoshi Son’s bet is characteristically aggressive, but today’s top funds are all at the table for fundraises of hundreds of millions and valuations of billions, with “a zero added to the size of the chip,” making SoftBank’s large checks no longer unique.
Oracle remains the clearer picks-and-shovels seller in this expansion, because it does not need to pick the model that ultimately wins. As long as data centers continue to be built and operated by Oracle, “I earn this service revenue and this money”; the same logic runs down the chain to Vertiv, liquid cooling, transformers, and cables. 刘一鸣’s sense was that if Stargate truly advances to Phase 5, none of the existing technologies seems capable of delivering it; 庄明浩 did not further confirm this engineering judgment in the conversation.
Even if GPT-5 launched as expected at the time, it was more likely to reinforce the existing paradigm than recreate a once-in-a-generation shock. Foundation models, reinforcement-learning reasoning, fusion, and MoE are already highly visible paths; 庄明浩 expected it to be “very strong,” rank highly, and score well, but perhaps not particularly surprising. If Scaling Law progress has indeed slowed, training compute would face more direct pressure. By contrast, he is more focused on whether DeepSeek V4 can raise the quality of the foundation-model base and further unlock the R-series reasoning models.
OpenAI’s monetization efforts are multiplying, but still lag far behind its infrastructure-spending curve. ChatGPT’s daily message count is said to have risen from roughly 1B at the start of the year to 2.5B, and its full-year revenue target was reportedly raised from $10B to $13B; but the Oracle contract alone is worth $30B a year, and total costs go far beyond that. “When will those two curves change” remains the $600B question Sequoia posed last year—and today that figure may need to double.