1/ the gap between 90% and 99.99% AI adoption is bigger than the gap between 0% and 90%.
2/ at 99.99%, agents don't wait to be called. they're already in the Slack thread, already surfacing what the team needs before anyone asks. not a tool. a coworker.
3/ agent-centric means agents are the default, not the exception. most orgs ask: where can AI help? an agent-centric org asks: where do we actually need a human? that inversion is the whole thing.
329
Zhipu AI IPO: Deciphering the Commercial Playbook of China’s "Anthropic" 🦄
Zhipu AI, China’s largest independent LLM unicorn and creator of the GLM series, has officially passed its HKEX listing hearing. This IPO isn't just a capital market milestone—it serves as a real-world case study on how Large Language Models (LLMs) are actually being commercialized in the Chinese market.
After digging through their prospectus, here are my core takeaways:
1. Business Model: Enterprise First, On-Prem is the Cash Cow; Contrary to the API-first strategies often seen in the West, Zhipu operates on a unique revenue mix. Its MaaS (Model-as-a-Service) approach stands on two legs:
- Localized Deployment (On-premise): Currently the revenue pillar, contributing ~85% of total revenue. It caters heavily to data-sensitive giants in finance, energy, and government sectors.
- Cloud Deployment: While currently only ~15%, this segment is critical for future scalability.
2. Financials: The Growth Logic of an Enterprise Software Vendor
Revenue Velocity:
- We see rapid growth from RMB 57.4M (2022) to RMB 312M (2024). H1 2025 alone recorded RMB 191M, a 325% YoY increase.
- Margin Analysis: Overall gross margin remains healthy at 50-60%.
Insight: On-prem margins are significantly higher than Cloud. As the cloud business expands, rising server and inference costs will likely pressure overall margins—a necessary growing pain for scale.
- Customer Concentration: The top 5 customers contributed 40% of revenue, though this concentration is trending downward as the client base diversifies.
3. R&D Costs: From "Talent War" to "Compute Arms Race" Zhipu’s 2024 R&D spend hit a massive RMB 2.96B—over 7x its revenue. However, the structural shift in where that money goes is the real story:
- Early Days: Costs were driven by people. In 2022, personnel accounted for 49.8% of R&D. By 2024, that share dropped to 16.7%.
- Now: It’s all about compute. Compute service fees have become the single largest expense, exceeding 70% in 2024. The nature of competition has shifted from a battle for brains to a capital-intensive war for GPUs.
Zhipu presents a unique hybrid profile: it generates revenue like a pragmatic enterprise software vendor but bears the heavy compute costs of a deep-tech giant. The key post-IPO question is whether this "software-integrated" commercial model can sustain the exponential costs of the ongoing compute arms race.
286
NeurIPS trip in San Diego was really amazing!!!! The scale and concentration of researchers offered a rare window into where the field’s micro-consensus is forming. Here are several insights I walked away with(open.substack.com/pub/hb1027…):
1. Scaling still works, but Smart Scaling is the new frontier
Linear and superlinear gains from more data and compute remain real. Yet smarter levers in post-training, such as higher-quality synthetic data in SFT or better exploration in RL, are now enabling smaller models to approach large-model capabilities.
2. Post-training is where the action is
Continual learning, reward shaping and new alignment techniques are drawing the densest research energy. Pretraining directions like diffusion-based LLMs, world models and RL-enhanced training exist but are early. Near-term progress will come from making models more usable, reliable and deployable rather than from pushing raw scale.
3. Data is the real bottleneck, not compute
High-quality human data is finite. Synthetic data is promising but inaccessible for most teams since high fidelity depends on strong teacher models. Approaches such as Prismatic Synthesis, which use mid-tier teachers to generate diverse and controllable data, show how the bottleneck may shift from scraping data to producing it.
4. RL and Agents still lack abstraction and active learning
Generalization is possible, but abstraction remains the missing piece. Today’s RL systems optimize inside human-defined state and time spaces, which limits agency. A true agent must invent and revise its own abstractions. In the near term, narrow vertical agents and deeper integration with foundation model post-training will be more viable than generic “agent platforms.”
5. LLMs are not the only path to AGI
Work at the intersection of AI and neuroscience is gaining legitimacy. The Sejnowski-Hinton Prize reflects a shift toward exploring learning rules and architectures that remain scalable under real-world constraints. Memory, OOD generalization and concept formation may all be projections of the same underlying paradigm shift.
6. Innovation requires alignment across academia, industry and capital
Breakthroughs with industrial impact rarely come from academia alone. They require synchronized progress across researchers, corporate labs, engineering teams and capital. The Bay Area’s unique density compresses this loop and explains why the frontier keeps emerging there.
A more detailed version is on my blog for anyone who wants to dig deeper into the investment angles. And it was genuinely energizing to meet so many new and familiar faces at NeurIPS — and fun to even run into @lexfridman for a quick photo.
119
I’ll be at NeurIPS (Dec 3–5)!
Been deep diving into the whole data layer lately — synthetic data, domain data engines, eval pipelines… The more I look, the more this feels like one of the biggest opportunities in AI right now.
If you’re a data-focused researcher or founder (or thinking about starting something), let’s meet and jam on ideas.
DMs open — excited to connect! 🚀
3
2
160