Summary

DeepSeek shifted the first half’s focus from “closed-source models leading” to a greater emphasis on open-source ecosystems, leading 张璐 to view Nvidia’s post-selloff outlook as bullish. Simpler architectures and training paths that do not depend on the most advanced GPUs could broaden AI adoption; the capital markets sent the opposite short-term signal, but she sees innovation moving from purely corporate-driven toward increasingly community-driven.
The host described AI agents as the next general-purpose platform after the PC and the internet; 张璐 emphasized that a true agent must handle complex tasks, make autonomous decisions, and choose its own tools. Coding agents such as Cursor and Windsurf depend on Claude, GPT, and Gemini, which means native products such as Claude Code could compress their moats; OpenAI once proposed acquiring Windsurf for $3B. The host argued that Google’s roughly $2.4B deal for the team, technology license, and know-how showed that, in Big Tech’s eyes, “money is cheap; time is the most valuable thing.”
张璐 identified OpenAI as the AI company facing the greatest challenges this year: public-facing C-end data is nearly exhausted, while the next step requires high-quality industry data and a direct fight with Microsoft in the to-B market. Microsoft remains constrained in the short term by its agreements and Azure ties, but it has already added models from Mistral and Cohere, strengthened internal development, and may allow Copilot to call multiple models dynamically; she characterized this as proactively preparing for “de-OpenAI-ization.”
Google’s AI technology and full-stack cost structure may be undervalued by the market, while Meta and Apple face, respectively, an ecosystem gap and a missing product. Google has TPUs, models, cloud, infra, and applications, but must decide when to dismantle search advertising, its “cash cow”; Meta is using cash, compute, and talent to “buy time,” and 张璐 remains bullish over the long term but does not think it can catch up by year-end, while Apple still has a hardware-entry window that is closing.
The talent war is not for ordinary engineering execution but for the “brains” capable of defining model architectures and product paths in the unknown, a group that may number no more than a few thousand people globally. Reported individual offers of $100M, Mira Murati’s company raising $2B in its first round, and Ilya Sutskever’s company raising $1B at a $5B valuation all reflect scarcity premiums; but 张璐 worries that first-round valuations have already priced out returns because “it’s not as if we’re going to have many trillion-dollar companies.”
AI is increasing capital efficiency while forcing large VCs to evolve from single-product venture firms into multiproduct financial institutions. Companies that once took three to five years to reach $2M-$3M in revenue can now reach tens of millions in a year; later-stage financing needs may fall from hundreds of millions of dollars to tens of millions, prompting large funds to expand deployment through RIAs, fund-of-funds vehicles, and other asset-management products.
M&A has become the primary liquidity repair mechanism for venture capital ahead of IPOs, but talent-driven deals may not allow investors to share fully in the value created. Fusion Fund had 4 exits in the first half, 3 involving important contributors to the open-source ecosystem; conventional strategic acquisitions can rapidly recycle capital and talent, while talent acquisitions often direct most consideration to the team, leaving VCs with relatively limited returns in Windsurf-style deals.
The U.S. IPO market cannot yet be described as reopened; the opportunities generating real revenue are concentrated more in vertical to-B, healthcare, and industrial AI. More than 20 Silicon Valley companies preparing to go public in the first quarter largely put their plans on hold, while CoreWeave and Circle remain only a handful of examples; meanwhile, U.S. healthcare accounts for roughly 20% of GDP, more than 30% of human data is healthcare-related, and less than 5% is being used, making this “data gold mine,” together with industrial and space automation, the deeper value of AI’s “shovels.”

Deep dive

1. DeepSeek Turned Open Source from an Alternative into the Industry’s Main Track

张璐 described the first half of 2025 as a period that “cannot be measured in months,” because models, products, papers, and deals were being refreshed almost every week; DeepSeek was the clearest cognitive turning point of the year’s opening months.
The day after DeepSeek was released, she happened to visit Nvidia. When the stock plunged on Monday, her conclusion, together with Nvidia’s strategic investment team, was bullish: the open-source ecosystem would drive broader AI applications and demand.
The more important signal was not simply whether performance could catch up with closed-source models, but that model architectures could become simpler and training might not require the most advanced GPUs. She also left room for another possibility: “Some new model architectures may run more efficiently on CPUs than on GPUs.”

2. Big Tech and Startups Are Moving from Competition toward Ecosystem Coexistence

张璐 has worked with Nvidia since 2017. This year’s GTC Inception Program already covered thousands of companies, while the new DGX Cloud Program offered selected companies backed by partner funds a certain amount of free compute and targeted engineering support.
Partners included major institutions such as Insight Partners, a16z, General Catalyst, and Lux, with Fusion Fund the only early-stage fund among them. This shows that Nvidia is also building the entrepreneur ecosystem and platform partnerships systematically.
Amazon, Google, and Microsoft are likewise attracting startups with free credits, cloud resources, APIs, and engineering support. 张璐’s summary is that large companies and entrepreneurs “have a competitive relationship, but even more, a collaborative one.”
Fusion Fund had 4 exits in the first half, 3 involving important contributors to the open-source ecosystem who were acquired by companies including Nvidia and Apple. This reinforced her view that AI innovation is shifting from corporate-driven to “community-driven.”

3. The Barrier to AI Agents Is Autonomous Decision-Making, Not Renaming Automation

The host described AI agents as “the next general-purpose platform after the PC and the internet.” 张璐 noted that agents are not a new concept—they existed more than a decade ago—but this wave has given them broad recognition.
Her minimum definition is demanding: an agent must handle complex tasks, make decisions without human intervention, and determine which tools to call from a given tool library. “Only when your product reaches this level can you call it an AI agent.”
Simple workflows, single-task automation, and ordinary digitization tools cannot all be folded into the agent category. The host mentioned Manus receiving investment from Benchmark; 张璐 confirmed that Manus had secured backing from Silicon Valley venture capital and was deploying into global markets, while also noting that China’s open-source ecosystem is influencing the global ecosystem.

4. Native Model Companies Entering the Market Has Suddenly Thinned the Moats of General-Purpose Coding Agents

Cursor and Windsurf had both become highly popular, reflecting the heat around coding agents; but both still call models such as GPT, Claude, and Gemini underneath. The question is therefore unavoidable: once model companies build coding products themselves, what advantages remain at the application layer?
After Claude Code launched, many entrepreneurs and companies around 张璐 began migrating because its native model, development environment, and fit with engineers’ habits were superior. It has not completely displaced Cursor or Windsurf, but it is already a directional warning.
The Windsurf deal unfolded like a “soap opera”: OpenAI once proposed acquiring it for $3B, after which Google pursued the team for roughly $2.4B, and Cognition later took the remaining team. Microsoft’s technology-sharing terms were one reason the OpenAI transaction failed, but not the only one.
The host argued that Google spent roughly $2.4B to obtain the team, technology license, and know-how directly while avoiding an antitrust investigation. 张璐 agreed with the core logic: “Money is cheap; time is the most valuable thing.”

5. OpenAI’s Real Bottleneck Is the Simultaneous Collision of Data, the To-B Path, and Microsoft

张璐 believes public-domain C-end data is “basically used up.” Further model improvement requires high-quality industry data, which is not public and can only be obtained through enterprise partnerships.
This forces OpenAI into the to-B market where Microsoft is strongest. Microsoft originally wanted OpenAI to supply the underlying models while it handled products and integration, but OpenAI is now building enterprise offerings, agents, and full-stack products, creating a direct clash in strategy.
Anthropic’s to-B path is relatively smooth. Microsoft, meanwhile, can promote Copilot through its existing sales, support, customer network, and Azure credits; OpenAI’s B-end path is more difficult because it must compete for business with its largest shareholder and execution partner.
When OpenAI delayed its open-source model and attributed the decision to potential security risks, 张璐 found the reversal striking: “OpenAI used to rarely cite safety as a central reason, and now it has started to emphasize safety.” Anthropic was originally founded because its founders believed OpenAI did not take safety and ethics seriously enough.

6. Microsoft Will Not Untether Immediately, but It Is Already Preparing for “De-OpenAI-ization”

Regulators are still examining whether Microsoft effectively controls OpenAI and whether monopoly concerns could arise, prompting Microsoft to give up its board observer seat. Sam Altman’s push to adjust OpenAI’s equity and corporate structure also produced only a compromise outcome.
Microsoft is unlikely to split completely in the short term because its existing agreements and Azure ties remain tight. But it is already supporting models from Mistral, Cohere, and others, accelerating model development at Microsoft Research, and reducing product dependence on a single supplier.
张璐 expects Copilot eventually to call multiple models dynamically rather than rely entirely on GPT. Enterprise customers’ security and compliance requirements also make multi-model supply a more rational form of insurance. This is gradual estrangement under “deeply intertwined interests.”

7. Meta Is Using Cash, Compute, and Talent to Compress the Time Lost on Strategy

Meta was once an open-source leader, but Llama 4 underperformed and DeepSeek weakened its leadership narrative. More damagingly, it had neither leading model capabilities nor a formed ecosystem.
The deal with Scale AI mainly addressed gaps in the data and tooling layers. According to 张璐, only 3 or 4 Scale AI employees actually entered Meta. Meta has also recruited other scarce talent aggressively, making the overall effort look more like a traditional attempt to catch up by stacking cash, compute, data, and people.
Meta is not short of engineering capability; it lacks the “brains” for model architecture and ecosystem building. Alexander Wang is valued not only for his technical ability but also for ecosystem construction, while upstream research talent is needed to help Meta assess unknown directions rather than execute an already-written roadmap.
张璐’s cautious conclusion is that new talent and the old organization will inevitably go through adjustment and growing pains, with no guarantee that newcomers will immediately devote themselves fully to products. “I’m still very bullish over the long term,” but she believes catching up by the end of 2025 will be “quite difficult.”

8. Meta’s Internal Problem Is Not Insufficient Effort but Poorly Connected Direction and Knowledge

The unconfirmed information she had heard was that, during Llama’s iterations, the next-generation team might not have access to all the information from the previous generation. In addition, the chief AI scientist had long been unconvinced by the large-language-model direction, leaving internal resource allocation misaligned.
Zuckerberg is making an almost “burn-the-boats” move, hoping the new team will create a catfish effect. Meta already has AI hardware entry points such as headsets and glasses; if it can fill the gaps in models and systems, its hardware advantage could become a complete ecosystem.

9. The AI Talent War Is for the “Brains” That Can Define Unknown Paths

Zuckerberg’s description of the shift in recruiting is telling: researchers used to ask how large their scope would be, but now care more about “having as few people report to me as possible while giving me as many GPUs as possible.” The scarce resource is not just compensation but research leverage.
张璐 estimates that people capable of defining model architectures, organizing paths involving LLMs and RL, and judging new CPU or GPU architectures may number “no more than a few thousand.” The most critical cohort is nowhere near large enough to be divided among all the major technology companies.
These companies already have engineers, infrastructure, and execution capabilities. What they are recruiting is not someone to implement a fixed concept, but someone who can answer: “You give me a concept, you give me an architecture—how are we going to build it?”

10. Grok Proved That More Force Can Produce Miracles and Made xAI the Second Half’s Wild Card

Grok went from Grok 1 in early 2024 to Grok 4 in just over a year. 张璐 believes some of its metrics have reached leading levels and called the progress “very, very astonishing.” It reflects a combination of resources, a brain trust, architecture, and execution.
She did not sidestep the controversy: Twitter data may introduce bias, and the model has produced Hitler-like conversations. But judged purely on other parameters, Grok 4 remains “extremely impressive.”
Roughly 70% to 80% of xAI’s internal code is already written by its own AI coding tools, although those tools have not yet been turned into a third-party product. 张璐 is looking forward to comparing them with Gemini and Claude Code once released, and sees Musk as the second half’s most unpredictable “wild card.”
This level of execution is not marketing. The host said he knew some of xAI’s founding members, who told him that Musk would work with them from 1 a.m. to 6 a.m., schedule another meeting for 7 a.m. or 8 a.m., sleep for a few hours, and continue working.

11. Google Leads in Technical Capability, but Search Advertising Prevents an All-Out Charge

Gemini 2.5’s long context window, small edge models, open-source Gemini CLI, and coding capabilities have all improved in succession. 张璐 also singled out AlphaFold and AlphaGenome’s applications to sequencing data.
The host noted that the head of Google Labs, which owns NotebookLM, also oversees Labs and the Gemini application. 张璐 confirmed that NotebookLM was built jointly by DeepMind and Google Labs, and believes Google is further integrating previously dispersed research and product divisions to convert R&D depth into product velocity.
张璐 personally ranks Google first among the major technology companies in AI capability and believes the market undervalues it. TPUs, models, cloud, infra, and applications form a complete full stack, turning model competition into the type of cost competition Google is best at.
The awkward part is that Google is fully capable of building a generative search product better than Perplexity, but cannot kill search advertising, its cash cow, before a new revenue model emerges. Younger users are using traditional search less and less, and changes in external habits will not give Google unlimited time.

12. Apple Still Has a Hardware Entry Point but Has Missed the First Round of AI Product Competition

张璐 had previously been very bullish on Apple’s AI chips: they are powerful and deeply integrated with its hardware ecosystem. The problem is that “it has no AI product”; powerful chips and devices have not translated into models or a tangible AI experience.
She described the deep integration of iOS 26 and ChatGPT as a “temporary compromise”: Apple is filling the gap through a partnership first and building out gradually, but other companies’ iteration speeds have exceeded its original expectations.
In the short term, agents will still most likely run on phones and computers, leaving Apple with a platform entry point. Over the long term, Meta, OpenAI, and Google are all exploring new hardware interfaces, and the window “may not be very long.”
Apple’s more realistic path may be to embed AI into headphones, chips, local models, and the operating experience rather than build an open model. Tim Cook is highly capable, but a professional manager is unlikely to wager his position on a potentially failed, outsized bet in the way a founder might.

13. Founder-Led Old Giants Are More Willing to Reinvent Themselves

The host cited Oracle and Salesforce to ask why companies from the previous generation could also capture AI. 张璐 observed that the fastest transformers are usually still led by their founders, allowing decisions to be more “steady, precise, and forceful.”
Salesforce has an Einstein model. Using Nvidia as an example, 张璐 explained that pairing proprietary models with proprietary applications and AI agents enables system-level vertical optimization, improving specialization, flexibility, and cost.
Hallucinations remain the central constraint on enterprise agents. Proprietary models, applications, and system-level optimization can help reduce hallucinations and meet B-end reliability requirements.

14. AWS Can Share in Growth by “Selling Shovels” Without Fighting for the Front Stage

AWS occupies the position of the shovel seller in a gold rush: model training, inference, and agent execution all consume cloud resources, so the more model and agent companies there are, the greater its room for customer and revenue growth.
Amazon is also an important supporter and investor in Anthropic. AWS’s free cloud credits can be used for Anthropic services, creating strategic coordination between the two. It may appear not to have launched a high-profile unified AI product, but in practice it can quietly win through infrastructure.

15. Super-Talent Financing Is Betting on Trillion-Dollar Companies, but First-Round Valuations May Already Have Consumed the Returns

The capital examples cited on the program were extreme: Mira Murati’s company raised roughly $2B in its first round, reportedly alongside an acquisition indication of roughly $10B from a technology company; Ilya Sutskever’s company raised $1B at a valuation of roughly $5B.
The capital is betting that these teams can build foundational general-purpose models and become trillion-dollar companies. 张璐 acknowledged that possibility, but believes the odds of a new company becoming the dominant general-purpose model provider are lower than several years ago because Google and other Big Tech companies have advanced so quickly, while the vertical-small-model path has yet to establish a clear winner.
A $100M individual recruiting price can explain the premium on team assets, but it cannot automatically prove investment returns. She has not invested in such companies because “it’s not as if we’re going to have many trillion-dollar companies,” and the growth cycle for a trillion-dollar company may not be short enough.

16. AI Is Increasing Capital Efficiency and Forcing Large VCs to Become Multiproduct Financial Institutions

The program noted that Lightspeed had newly registered an RIA, while a16z, Sequoia, and General Catalyst already had similar arrangements. 张璐 believes large funds are shifting from single-product VCs toward comprehensive financial institutions combining fund-of-funds vehicles, asset management, and other products.
One fundamental reason is that there is too much money: once fund sizes reach several billion or even more than $10B, traditional venture stages alone make effective deployment difficult. Funds must lengthen investment horizons and broaden asset classes.
AI is also rewriting the pace of company financing. In the past, it might have taken 3 to 5 years to go from zero to $2M-$3M in revenue; now that can happen in a year, with some companies reaching tens of millions. A company that once raised $10M to reach $5M in revenue might now raise $5M and reach a larger scale.
When later-stage funding needs fall from hundreds of millions of dollars to tens of millions, large funds have an even harder time deploying sufficient principal. The RIA shift is both an expansion strategy and a response to startups’ declining capital requirements.

17. M&A Repairs Liquidity While Reordering the Returns of Teams and Investors

Strategic buyers do not price acquisitions solely on the target’s existing revenue. If integrating a mature product into their own channels and customer base can immediately create $500M or $1B in value, paying an acquisition price of the same magnitude can still be rational, with future products providing additional optionality.
Fast acquisitions recycle capital and talent into the innovation ecosystem and eliminate the need for every company to reach an IPO. For VCs that have endured several years of weak exits, this is a faster capital-return mechanism than waiting for public markets.
Talent acquisitions, however, often direct more consideration to founders and teams, compressing investor returns. Even if the Windsurf deal is large, its investors may still receive relatively little, forcing VCs to redesign terms and reassess what kind of exit qualifies as a “win-win.”

18. CoreWeave and Circle Have Performed Well, but Still Do Not Prove the IPO Market Has Reopened

The market initially expected an IPO recovery. More than 20 Silicon Valley companies were preparing to list in the first quarter, and several Fusion Fund companies were also preparing; after tariffs and financial-market volatility, those plans “basically all went on hold.”
张璐 directly rejected the claim that “the IPO market has reopened”: CoreWeave and Circle performed well after listing, but the sample is too small to represent a restored channel.
Stripe and Databricks have been awaited for years, but their strong cash flow and financial condition allow them to choose their timing. She offered a contrarian interpretation: the ability to pause an IPO because market conditions are poor may itself show that a company has not been forced by cash flow to list “even while bleeding.”
Her view of the second half is conditionally optimistic: if political and policy uncertainty becomes the norm, Wall Street’s shock response may diminish and companies may become more willing to list. But as of the interview, “the IPO market remains in a state where it has not reopened.”

19. AI Delivering Real Commercial Value Is Moving Deeper into Healthcare, Industry, and Space

Over a decade, Fusion Fund has invested in more than 100 companies, concentrated in enterprise AI, industrial automation, and healthcare. Its infrastructure portfolio includes Lepton, Voyage AI, You.com, and Vectara, which has stood out in RAG and embedding models; the companies forming revenue quickly are more often vertical to-B agents than the image and text applications most visible in the media.
Healthcare is the largest “data gold mine”: it accounts for roughly 20% of U.S. GDP, more than 30% of human data is healthcare-related, yet less than 5% is currently being used. Portfolio examples include a vertical cell-therapy model backed with Khosla Ventures; she also mentioned the Evo 2 sequencing model released by Arc Institute and Subtle Medical’s generative imaging enhancement.
Subtle Medical has received multiple FDA approvals and can enhance low-resolution CT or MRI scans taken in minutes to high-quality images, reducing time, cost, and radiation exposure. 张璐 emphasized that AI’s role is to “empower” doctors, nurses, and underlying medical technologies, not simply replace practitioners.
In industry, supply-chain realignment across North America is accelerating demand for automation and robotics. The space industry has integrated 3D printing, AI, and robotics from the outset. The lunar robotic system she cited can already envision extracting water and other supplies on the Moon to serve as a refueling station for interstellar travel, while the moon rover has reached version 11.
This ultimately returns to her “shovels” metaphor: AI is not merely improving business models; it is raising baseline productivity in healthcare, manufacturing, and space. Technological development carries risks, but stagnation may do more damage. Using AI tools will eventually be like using a computer—not a differentiating skill, but a basic requirement.

Vol. 69: Revisiting the Stargate Project — with 刘一鸣/101 Weekly 20Product: Should We Kill the PM Role & How Monzo Builds Product