DeepSeek Challenges Nvidia’s Software Lead in China

Caught My Eye…

1) DeepSeek Challenges Nvidia’s Software Lead in China

On September 30, DeepSeek open-sourced a set of programming tools for Huawei’s Ascend chips. One of them, DeepGEMM-Ascend, lets developers use similar commands to DeepSeek’s Nvidia software, making it easier to move AI workloads from Nvidia chips to Huawei’s.

That matters because software is a big part of Nvidia’s advantage. More than 7.5 million developers use CUDA and Nvidia’s related tools, and much of their code is built around that ecosystem. DeepSeek’s new tools reduce some of the work required to switch hardware.

DeepSeek reportedly plans to deploy at least 160,000 Huawei Ascend 950DT chips to run its models and answer user queries while still training its models on existing Nvidia chips. Deepseek founder Liang Wenfeng reportedly told investors he plans to use more Huawei chips for training too.

Huawei still has two major constraints. The first is supply. Component shortages are expected to limit Ascend 950DT production to roughly 200,000 to 300,000 chips this year, meaning DeepSeek’s planned deployment alone would account for roughly 53% to 80% of annual output.

Performance is the other. Liang reportedly told investors that a model as large as today's biggest would take about 200,000 Huawei chips against 50,000 Nvidia GB300s, his rule of four Huawei chips for every Nvidia chip, with Huawei about two years behind.

Huawei reportedly plans to sell these chips only in China for now, so the near-term impact on Nvidia is largely confined to the Chinese market. Nvidia is already assuming no data-center computing revenue from China this quarter.

2) Gemini 4 Argon and the New Release Trade-Off

Google announced Gemini 4 Argon on September 30, with initial external access limited to vetted cybersecurity defenders and trusted testers. Paying developers and Google AI Ultra subscribers are next, but there is currently no public release date. Google says the model can find and fix serious software vulnerabilities on its own, a capability that could also help attackers. Trusted defenders will receive a version without restrictions on cybersecurity tasks while Google strengthens safeguards for broader access.

In Google's comparison table, Argon leads or ties on 14 of 19 benchmark results against the best models from OpenAI and Anthropic. The widest margin came from Harvey's test of legal research and drafting, with Argon at 19.6%, against 6.7% for the nearest rival.

The introductory price is $2 per million input tokens and $10 per million output tokens. That is a fifth of OpenAI's rate for GPT-6 Astra ($10 and $50) and half of Anthropic's for Claude Opus 5.5 ($4 and $20). After the introductory period, Argon will cost roughly the same as Opus.

Two days earlier, OpenAI shelved the planned release of GPT-6.1 Astra after it failed the company's safety bar. Saachi Jain, OpenAI's head of safety systems, described a trade-off between greater task persistence and unauthorized behavior. Google's rollout adds another example of the gap between what an AI lab can demonstrate and what customers can use.

3) Washington Opens Up the Credit Score Business

On September 28, the regulator for Fannie Mae and Freddie Mac said the two would eliminate a pricing adjustment that effectively treated VantageScore as 20 points lower than FICO. Under the new system, a 720 VantageScore and a 720 FICO score will fall into the same pricing tier. Shares of Fair Isaac, which makes the FICO score, fell 26.5% the next day.

Three credit bureaus, Equifax, Experian and TransUnion, keep records of how Americans borrow and repay. A scoring formula turns a record into one number, usually 300 to 850, where higher is better. Fair Isaac owns no records. It licenses its formula to the bureaus and is paid for each score they sell. VantageScore is a rival formula the bureaus created together in 2006.

For decades, Fannie and Freddie, the government-controlled companies that buy mortgages, required a FICO score on every loan whenever the borrower had one. This year they began letting lenders use VantageScore instead. Fair Isaac's standard price this year is $10 a score, though it says lenders were already paying about that through the bureaus. VantageScore costs $0.99 at TransUnion.

Fannie and Freddie charge a fee per loan that falls as the borrower's score rises. VantageScore often gives the same person a higher number than FICO, so the price table had counted every VantageScore as 20 points lower. That offset is now gone. Take a borrower with a 700 FICO score but a 720 VantageScore. Under the old pricing, lenders effectively treated that 720 VantageScore like a 700. Under the new system, lenders can use the full 720 score, which cuts the fee from 1.5% of the loan to 1.25%, or $1,000 on a $400,000 mortgage. Rocket Mortgage, America’s largest retail mortgage lender, said it would make VantageScore its default.

Fair Isaac CEO Will Lansing argues that letting lenders choose between FICO and VantageScore encourages “gaming.” If the same borrower gets a 700 from FICO and a 720 from VantageScore, a lender could use the higher score to get better pricing.

4) Micron's Record Profits

Micron, which makes the memory chips that store data for processors, reported $54.2 billion in quarterly revenue on September 30, up from $11.3 billion a year earlier. Net income reached $37.7 billion. Its gross margin, the share of sales left after production costs, was 86.8%, above Nvidia's 75.0% in its latest quarter. AI systems need fast memory besides their processors to keep them supplied with data. That demand has helped lift profits in a business known for sharp earnings swings.

Investors still value those earnings differently. As of this week, StockAnalysis put Micron at ~6x expected annual earnings, against ~19x for Nvidia. Memory has a history of shortages followed by gluts: manufacturers build factories when prices are high, then new supply drives prices down. Micron lost $5.8 billion in fiscal 2023. A low multiple can reflect concern that today's earnings will fall, even when the latest quarter looks exceptional.

Micron revealed it now has 26 multi-year Strategic Customer Agreements (SCAs). These cover over 35% of its entire expected revenue through 2030, backed by $32 billion in upfront customer commitments and cash deposits. They require customers to take the chips or pay for the contracted supply. The price floors offer protection when chip prices fall, while the ceilings limit gains when prices rise. Ultimately, this pricing-band strategy structurally alters Micron's risk profile. By securing a guaranteed revenue floor that management notes is well above any prior cyclical peak margin, Micron mitigates the downside of a market oversupply.

One of our upcoming Deep Dives explores the memory ecosystem and examines how durable the demand is. Subscribe to get access.

5) AI Beats 12 Licensed Accountants

Mercor, a company that pays experts to build training data and tests for AI labs, put 12 licensed accountants against AI on four month-end tasks. The accountants, who averaged about 5.5 years of experience, had to search through a simulated company’s files, find the right numbers, do the math, and fill in a results table.

On average, the accountants got 37% of the required answers and calculations right, even though most finished before the three-hour limit. Claude Opus 5 got every required item right on all 20 attempts, each in under 10 minutes.

The four tasks were simplified versions of Mercor’s much harder accounting benchmark, which was designed to expose where AI still fails. They stacked hard-to-spot but realistic errors, and on the full test, AI is far from perfect. Across 160 harder tasks from 10 simulated companies, the best model scores about 62%.

When Mercor launched that test in July, it ran each model on every task eight times, and even the most consistent one got only 2.6% of tasks right on all eight runs. Most failures came from flawed reasoning, such as spotting an error early and leaving it out of the final entry. In the new study, accountants working with Claude took about 15 times as long as Claude alone and also scored slightly lower.

Mercor puts the model's cost at no more than $0.21 for each required item it got right, against $10.35 for an accountant at the U.S. median wage, and expects accountants' work to shift toward clients and judgment calls.