DeepSeek this gun is already pressed against the back of Silicon Valley's head

DeepSeek V4 Released, a Few Days Later, Silicon Valley Continues to Stir, Bringing Up Several More Sharp Questions: Model Efficiency, Chip Landscape, IPO Timing, Open Source vs. Closed Source.

On April 29th, a video blog from Silicon Valley 101 invited chip architect Xiao Zhibin and former OpenAI researcher Jenny Xiao, who discussed the whole story in over an hour, thoroughly.

Surpassed by Open Source, Your Business Value Becomes Zero

The harshest statement in the discussion was a concept Jenny proposed last year—kill line, the death line drawn by open-source models for closed-source foundational model companies.

“If you’re a foundation model company and you get surpassed by open source, the value of your business is essentially zero.”

If you’re a foundational model company and open source surpasses you, your business value drops to zero.

This isn’t just technical competition; it’s a matter of life and death.

Jenny gave an example with Anthropic: if one day Claude is no longer the best model for programming, who will still use Claude Code?

Following this logic, on April 24th, after DeepSeek V4 was released, all Silicon Valley closed-source companies’ stock prices and valuations faced a soul-searching question: Does your model still justify this price?

The most direct way to compare prices:

GPT-5.5 is twice as expensive as GPT-5.4, with the Pro version for long texts costing $180 per million tokens.

On the same day, DeepSeek V4 was released. Input 1 yuan per million tokens, output 24 yuan per million tokens. The Flash version is even more aggressive: input 0.2 yuan per million tokens, output 2 yuan per million tokens.

One is twice as expensive, the other ten times cheaper.

The valuation of foundational model companies is binary—your reason for existence is the strongest model. Once you’re no longer the best, your valuation drops to zero. Even if you’re OpenAI.

Too Much Money, Yet No Savings

Jenny exposed a harsh truth in the discussion that Silicon Valley dares not face:

“Silicon Valley companies have too much money, which actually reduces their motivation to optimize efficiency. Chinese model vendors, under resource constraints, entered token efficiency innovation earlier.”

Resource constraints, paradoxically, accelerate innovation.

From day one, OpenAI believed in “move fast, break things,” buying GPUs freely and building infrastructure wildly. Anthropic, on the other hand, restrained itself, fearing that revenue wouldn’t keep up and that procurement costs would kill it.

What was the result? Under the same revenue, Anthropic’s capital efficiency is significantly higher than OpenAI’s.

Even more troubling, OpenAI is simultaneously fighting on multiple fronts—hardware division, self-developed chips, shopping apps—yet the core ChatGPT experience isn’t well optimized. Since late last year, many side projects have been cut, including Sora.

Investors’ mindset has completely changed. Previously, they saw AI companies as “the curve is still exponentially growing, keep investing.” Now, the question is:

“If we invest another 1 billion, 10 billion, where’s the marginal benefit? ROI where?”

DeepSeek’s answer is straightforward: continuing to expand infrastructure ROI may no longer be cost-effective.

Innovation is driven by necessity. Cheapness itself is a prerequisite for technological revolution.

Every industrial revolution is driven not just by technological prowess but also by how affordable the technology becomes. Only when it is cheap enough for ordinary people to use can technology truly change the world.

Without efficiency, AGI is just a demo

Xiao Zhibin, after reading the V4 paper, believes: “The direction is expected, but the engineering completion is surprisingly.”

All technical optimizations in V4 are aimed at the same target—token efficiency.

It uses three main tools:

• Muon Optimizer: Replaces some training modules with traditional Adam, further speeding up convergence.

Together, these three point to the same result: lower cost per token to generate, less memory used per inference.

Computational costs are reduced to one-third of Silicon Valley models, and memory usage is only one-tenth.

But what truly makes V4 terrifying is far more than just “saving money.”

Jenny repeatedly emphasizes in the discussion: in the chatbot era, token consumption is limited; a slightly more expensive model is tolerable for users. But the agent era follows a completely different logic—long task decomposition, multi-tool invocation, reflection and planning—token consumption is 10 to 100 times that of chatbots.

If each token is costly, models can’t think deeply over long periods, nor serve users at scale.

So she made a defining statement:

Without efficiency, AGI can only be a demo. With efficiency, AGI can become a real product.

In the agent era, efficiency itself becomes part of intelligence.

Why does Anthropic surpass to 1 trillion? Focus > Doing Everything

Recently, Anthropic’s valuation surpassed OpenAI, reaching 1 trillion dollars.

Jenny listed three reasons, but fundamentally it boils down to five words: Focus > Doing Everything.

First, Claude Code.

Why is Claude Code Anthropic’s “defining moment”?

Anthropic’s models have always been good, but Claude Code is the product that truly drives revenue. Peter Steinberger, founder of OpenClaw, wrote an article: “Claude Code is my computer.”

Once the model can write code, it can perform general tasks—updating CRM, forwarding emails, automating workflows—all fundamentally code-based.

Jenny’s sharp judgment: Programming is the most important step toward AGI. Whoever masters programming, is likely to be the dominant player in the AGI era.

Second, corporate trust.

Enterprise clients in Jenny’s fund repeatedly say the same thing: they choose Anthropic because it offers safety commitments. Plus, the incident where Anthropic sued the Pentagon has conveyed a strong signal to enterprises.

Third, avoid doing unnecessary things.

OpenAI wanted to build “everything for everyone,” but ended up dispersing its efforts and losing technical leadership. Anthropic focuses on three lines: safety, enterprise, and programming.

Silicon Valley investors hold a simple principle: Prioritize enterprise revenue over consumer revenue. Anthropic’s revenue is highly concentrated in enterprise, which is exactly the story that the US capital market loves most.

Nvidia: Short-term Security, Long-term Differentiation in Reasoning Market

Regarding chips, the current consensus is that DeepSeek is “de-Nvidia-izing.”

But the reality is more nuanced.

Training phase: DeepSeek V4’s massive pretraining was undoubtedly done on large Nvidia clusters. The V4 technical report (pages 16 and 20) mentions TCGenO5 and MegaMoE², which are deeply tied to the CUDA ecosystem for low-level optimization.

Adaptation phase: Huawei Ascend promotes “zero-day adaptation for fine-tuning/inference,” AMD promotes “integration and optimization on ROCm.”

Note a key phrase—“adaptation.”

This means the model has already been trained and finalized on Nvidia clusters. Ascend and AMD are doing post-training “interface” work with their own software stacks. It’s backward compatibility, not native replacement.

From this perspective, in the short term, Nvidia’s training moat is deeper than many think. The CUDA ecosystem isn’t something that can be moved away in a year or two.

But in the long run? Inference market is indeed loosening.

After V4 reduces the cost of long-context attention, the threshold for large-scale inference drops significantly. Inference is no longer about “who has more GPUs wins,” but “who has the right architecture wins.” Google TPU, AMD, cloud providers’ self-developed chips, and even domestic computing power are all seeking opportunities along this line.

The “80/20” signal that keeps Silicon Valley awake

Jenny’s portfolio data shows: 80% of tasks run on small to medium open-source models. Only 20% of the most complex tasks use closed-source models.

A year ago, no one believed this ratio.

Now, Silicon Valley is daily seeing messages like: “We hold $10 million worth of OpenAI stock, are you funds buying? Or do you know anyone who is?”

On April 29th, in Silicon Valley 101, Jenny summed up the whole story in one sentence:

“DeepSeek is like a gun pressed against the back of Silicon Valley model companies. If they don’t move fast enough, DeepSeek will catch up and completely destroy their business.”

This gun is already cocked.