U.S.-China Relations
NextFin News - DeepSeek, the Chinese AI lab that shocked global markets last year by claiming it trained a frontier model for just $5.6 million, opened its first external funding round seeking at least $300 million at a valuation above $10 billion - and watched that number climb past $20 billion as Tencent and Alibaba fought for a seat at the table. The bidding war, reported across April and June 2026, is the fundraising frenzy of the headline. But it is happening against a darker backdrop: the shadow market for the restricted Nvidia chips that Chinese labs depend on has tripled in price, turning China's AI dream into an increasingly expensive underground operation.
The tension at the heart of this story is simple. DeepSeek built its global reputation on the claim that clever engineering could beat brute-force spending. Now the company is doing what it swore it would never do - taking outside money - while the very market that made its low-cost narrative possible is pricing itself out of reach. The question is not whether DeepSeek can still compete. It is whether the entire premise of Chinese AI efficiency was ever more than a one-time accounting trick.
The Fundraise: From Reluctance to Frenzy
According to people familiar with the matter, DeepSeek began discussions in April 2026 to raise a minimum of $300 million in primary capital at a pre-money valuation of at least $10 billion. The deal marked a sharp reversal for founder Liang Wenfeng, who had personally funded the Hangzhou lab since 2023 through his quantitative hedge fund, High-Flyer Capital Management, and had rebuffed approaches from prominent Chinese venture capital firms and major technology companies.
Then the dynamic changed. By late April, with initial discussions sparking intense interest, the valuation target had climbed above $20 billion, according to subsequent reporting. Tencent Holdings and Alibaba Group entered talks to invest, and by early June some people briefed on the matter put the round at roughly 50 billion yuan ($7 billion) with a post-money valuation between 350 billion and 400 billion yuan - $52 billion to $59 billion. Liang himself reportedly committed 20 billion yuan, with Tencent considering 10 billion yuan and battery maker CATL looking at 5 billion yuan. None of the companies commented, and the figures remain subject to change.
Even the conservative $10 billion starting point represented nearly a threefold jump from the $3.4 billion valuation DeepSeek carried on secondary markets in 2025. Yet at $20 billion - or $50 billion-plus - the company still trades at a steep discount to American frontier labs: OpenAI is marked at roughly $500 billion in recent secondary trades, and Anthropic sits near $350 billion following a rejected tender offer. High-Flyer is expected to remain the controlling shareholder. The stated purpose of the raise is to strengthen financial reserves, expand computing resources, and offer higher salaries to attract and retain talent - the three line items that define the capital intensity of modern AI.
The timing is the tell. High-Flyer posted a reported 56.6 percent return in 2025, generating enough cash flow to bankroll DeepSeek's training runs, salaries, and data center leases for roughly 30 months without dilution. That a hedge fund profit stream - however strong - has finally reached its limit says something about the burn rate of the next generation of reasoning models and agentic systems. Training a frontier model costs tens to hundreds of millions of dollars per run; serving it to consumers at sub-cent pricing requires continuous GPU purchases and data center expansion. Liang told Chinese state media earlier in 2026 that "China's AI industry needs confidence as much as capital," a framing that now reads less like a statement of strength than a justification for opening the cap table.
The $5.6 Million Mirage
To understand why the fundraise matters, you have to go back to the number that started it all. In its DeepSeek-V3 technical report published in December 2024, the company stated that training consumed 2.788 million GPU-hours on a cluster of 2,048 Nvidia H800 GPUs, at a total cost of $5.576 million - calculated at an assumed rental price of $2 per GPU-hour. The report broke the figure down: 2.664 million hours and $5.328 million for pre-training, 119,000 hours and $238,000 for context-length extension, and 5,000 hours and $10,000 for post-training. It also carried a caveat: the figure covered only the model's "official training" and excluded costs tied to "prior research and ablation experiments on architectures, algorithms, or data."
That caveat did the heavy lifting. The $5.6 million number was not the cost of building DeepSeek. It was the marginal cost of the final training run, amortized over a cluster the company already owned, with years of failed experiments written off elsewhere. It was the AI equivalent of quoting only the rental fee for a cross-country road trip while leaving the $250,000 Ferrari out of the story.
The market believed it anyway. On January 27, 2025, Nvidia shares plunged 17 percent, wiping out $589 billion in market capitalization - the largest single-day loss for any company in U.S. history. Investors read the $5.6 million figure as proof that the moat of AI capital expenditure was an illusion, and that Chinese ingenuity had found a back door through America's export controls. Nvidia itself said DeepSeek's advances were an example of what was possible "leveraging widely available models and compute that is fully export control compliant."
The reality was more complicated. A research firm estimated that DeepSeek's actual hardware spend was "well higher than $500 million." AI experts said they expected earlier development stages to have relied on a much larger quantity of chips, an investment that "could have cost north of $1 billion." A source from an AI company familiar with large-model training said "around 50,000 Nvidia chips" were likely used. Alexandr Wang, the chief executive of Scale AI, said in a Davos interview that DeepSeek had secretly acquired access to a pool of 50,000 Nvidia H100 GPUs - the most advanced chip, banned for export to China. DeepSeek has said it has access to 10,000 of Nvidia's older A100 GPUs, obtained before the United States imposed export controls.
The H800 chips DeepSeek trained V3 on are Nvidia's China-compliant variant of the H100, with chip-to-chip data transfer of roughly 300 gigabytes per second compared with the H100's 600 gigabytes per second. The H800 was slower, but it was legal, available, and cheap - at least until it wasn't.
The Shadow Market Squeeze
Here is where the fundraise meets the market. The cheap-GPU environment that made DeepSeek's narrative plausible is disappearing, and the mechanism is a pincer movement between Washington and Beijing.
On the American side, enforcement tightened at the end of 2025. On March 19, 2026, prosecutors charged Yih-Shyan "Wally" Liaw, a co-founder of server maker Super Micro Computer, along with two others, with conspiring to route at least $2.5 billion worth of servers containing Nvidia AI chips to China in violation of export laws. The indictment alleged the scheme moved servers with banned H200 and B200 GPUs through an unnamed Southeast Asian company to Chinese buyers, using thousands of fake servers with shipping labels affixed by hair dryers. Authorities in Taiwan and Malaysia subsequently opened their own smuggling investigations, drying up the re-export routes traders had relied on. Nvidia said building data centers from smuggled chips is a "dead-end," adding that it provides no support or repairs for restricted products.
On the Chinese side, Beijing closed the legal channels from the other direction. After the Trump administration approved exports of Nvidia's H200 chip to China, Chinese customs were instructed to block the chips at the border. Commerce Secretary Howard Lutnick later confirmed that Nvidia had not sold a single H200 to a Chinese company months afterward. Both moves funnel demand toward the same destination: Huawei, which launched its Ascend 950PR inference chip on March 20, 2026, with 750,000 units planned for 2026.
The price consequences have been severe. Servers built around Nvidia's five-year-old A100 accelerator have roughly tripled since late last year, climbing from about 200,000 yuan ($22,300) to as much as 600,000 yuan ($82,000), according to chip traders. Nvidia's restricted Blackwell hardware sits at the top of the same market: the DGX B300 system, which retails in the United States for nearly $400,000, now trades above 8 million yuan ($1.1 million) on China's black market - more than double in six months. The RTX 6000 Pro workstation card rose from roughly 50,000 yuan ($5,580) at the start of 2026 to as much as 130,000 yuan ($14,500).
Rentals are no escape. A survey of GPU rates found that prices inside China now match or exceed U.S. levels, reversing the discount that abundant smuggled supply once provided. Buyers are repurposing gaming GPUs and bidding up older Ampere hardware, while a DRAM and HBM shortage works its way through every tier of the AI hardware stack, compounding component costs.
"Building data centers from smuggled chips is a dead-end," Nvidia said - a warning that applies as much to DeepSeek's next training run as to any other Chinese lab.
The macro picture confirms the squeeze. Bernstein projects Nvidia's China AI chip market share falling from roughly 40 percent in 2025 to about 8 percent by the end of 2026, with Huawei's share rising to around 50 percent. TrendForce separately estimates domestic Chinese chips will hold close to 90 percent of the country's AI hardware market by year-end. The transition is not frictionless: Huawei's legacy CANN software stack has long trailed Nvidia's CUDA ecosystem, though its upgraded CANN Next platform now mimics CUDA's programming model - thread blocks, warps, kernel launches - to lower switching costs. Domestic supply still cannot fully absorb the demand the import freeze has redirected, but the gap is narrowing faster than Washington anticipated.
What the Fundraise Really Prices In
The valuation - whether $10 billion, $20 billion, or $50 billion-plus - is not a bet that DeepSeek has solved the cost problem. It is a bet that the company can survive it. Three readings of the number are possible, and they point in different directions.
The bullish reading: DeepSeek has proven that architectural efficiency - mixture-of-experts routing, FP8 mixed-precision training, multi-token prediction - can compress the cost of frontier training by an order of magnitude. If the $5.6 million figure was understated but the underlying engineering is real, then DeepSeek can iterate faster and cheaper than American rivals even in a sanctioned hardware environment. The gap between the top U.S. model and the best Chinese competitor stood at just 2.7 percentage points as of March 2026, according to Stanford University's annual AI Index report. At $20 billion, DeepSeek is a 25x discount to OpenAI for a model that is not 25x worse. For strategic buyers like Tencent and Alibaba, that is not expensive - it is a hedge on the one asset class they cannot afford to miss.
The bearish reading: the fundraise is a distress signal dressed as a growth round. The cheap-GPU window that produced DeepSeek's breakthrough is closed. Every subsequent training run will cost multiples of the last one, because the hardware must now be sourced through a shadow market that has tripled in price, or migrated to a domestic stack that still lags in mature tooling. The capital raised will be consumed by capex and salary inflation before the next model ships, and the valuation will look like the top of the cycle. The frenzy itself is the warning: when strategic buyers bid a private AI lab from $10 billion to $50 billion in eight weeks, they are not pricing fundamentals. They are pricing panic.
The structural reading, which I find most persuasive: DeepSeek's low-cost breakthrough was a one-time convergence of three conditions that cannot be replicated - a pre-existing stockpile of restricted chips acquired before the tightest controls, a hedge fund willing to absorb years of R&D losses, and a rental market flooded with smuggled supply that kept marginal training costs artificially low. Remove any one of those and the $5.6 million number evaporates. The company is not raising money to scale a proven advantage. It is raising money to rebuild its advantage under new, harsher constraints.
Second-Order Effects: Who Wins and Who Pays
The first-order effect of the fundraise is obvious: DeepSeek gets cash. The second-order effects are more interesting, and they cut across the AI stack.
For Nvidia, the irony is sharp. DeepSeek's success was supposed to be the bear case for the chip designer - proof that its hardware moat could be engineered around. Instead, the tightening of export controls has made Nvidia's restricted chips more valuable, not less, inside China. The black-market premium on A100 and Blackwell hardware is a backhanded endorsement of CUDA's indispensability. But the revenue never reaches Santa Clara: every transaction happens in the shadows, untaxed and unsupported, and the long-term trajectory points to Huawei. Nvidia's China AI chip share collapsing from 40 percent to 8 percent is not a cyclical dip; it is a market being walled off.
For Huawei, the windfall is real but conditional. A captive domestic market with import substitutes priced out of reach is the best demand environment a chipmaker could ask for - with ByteDance reportedly committing more than $5 billion in Ascend orders. Yet the conditionality matters: CANN Next narrows the software gap, but a CUDA-compatible translation layer is not the same as CUDA's two decades of optimized libraries. Huawei's window to lock in developers before import policy shifts again is measured in quarters, not years.
For the global GPU rental market, the distortion is a gift. The market is valued at roughly $52 billion in 2026, up from $34.6 billion in 2025, and is projected to reach about $199 billion by 2031 at a 30.7 percent compound annual growth rate. Chinese demand, denied legal supply, leaks into whatever channels remain open, keeping rates elevated even as new capacity comes online elsewhere. The bifurcation - cheap, abundant compute in the West; expensive, scarce compute in China - is now a durable feature of the industry, not a temporary shortage.
For DeepSeek's open-weight strategy, the calculus changes. Releasing models under a permissive license drove enterprise adoption across Asia and Europe precisely because it let users with chips run the models for free. But if chips are the scarce input, giving away the model is a way of subsidizing competitors who have better hardware access. The open-source mantle that distinguished DeepSeek from OpenAI becomes a liability when the bottleneck shifts from algorithms to silicon.
The Counter-Thesis
The strongest case against this reading is that it mistakes a financing decision for a strategic one. DeepSeek is not raising because it is running out of money; it is raising because a high valuation is a useful tool for talent retention and partnership leverage in a market where rivals are poaching researchers with equity packages no hedge fund can match. The round may be as much about setting a valuation benchmark and anchoring employee option grants as about funding the next training run.
There is truth to this. A round at a high valuation can signal credibility to enterprise customers without meaningfully diluting the founder, and strategic buyers see value beyond the balance sheet. But the counter-thesis still depends on the hardware question, and the hardware question is getting worse, not better. If the smuggling crackdown holds and domestic supply does not mature on schedule, no amount of valuation engineering changes the unit economics of the next training run. The fundraise buys time; it does not buy chips.
The falsifying signal is specific: if Chinese customs begin clearing approved Nvidia H200 shipments at scale within the next two quarters, or if Huawei's Ascend 950PR achieves performance parity with Nvidia's H100 on standard AI benchmarks while maintaining volume supply above 100,000 units per quarter, then the shadow-market squeeze thesis breaks down and DeepSeek's cost position improves materially. Watch the A100 server price in yuan - if it falls back toward 300,000 yuan from 600,000 yuan, the pressure has eased. Until then, the constraint holds.
What to Watch
In the short term, the closing of the round - its final size, its investor mix, and whether High-Flyer retains control - will set the tone for Chinese AI valuations. A round that clears at $20 billion or above would signal that capital is still available for Chinese AI despite geopolitical friction. A round that shrinks or stalls would confirm the discount is structural.
In the medium term, watch DeepSeek's next model release and whether the company revises its cost disclosures. If the next training run is priced at multiples of $5.6 million, the efficiency narrative is being quietly retired.
In the long term, the decisive variable is not DeepSeek at all. It is whether China can build a domestic AI hardware stack that does not depend on smuggled Nvidia silicon. Huawei's share trajectory toward 50 percent, the CANN Next maturation curve, and the HBM supply constraint are the three dials to watch. If all three move in Huawei's favor, DeepSeek's fundraise will be remembered as the moment Chinese AI grew up. If they don't, it will be remembered as the moment the bill came due.
DeepSeek proved that clever code can narrow the gap with brute-force spending. What it could not prove is that clever code can repeal the laws of semiconductor economics. The fundraising frenzy is the price of finding out which one matters more.
Explore more exclusive insights at nextfin.ai.
