South Korean startup DeepX plans to spend more than $100 million developing a next-generation AI chip, the DX-M2, on Samsung Electronics’ 2 nanometer manufacturing process, Bloomberg reported on 13 November. That is 7.5 times the about $13.3 million of purchase orders DeepX had secured by October, as TheElec reported from the company’s statement, by our arithmetic.
DeepX’s own DX-M2 page lists 80 TOPS of NPU performance, models of about 100 billion parameters and about 5 W of power. Against the 25 TOPS and 4 billion parameter models of its first chip, the DX-M1, that is 3.2 times the compute and 25 times the model size on a similar power budget, by our arithmetic.
Business Pill · On-Device AI
A short explainer on running AI models on the device itself rather than in a distant data center. It teaches the general idea only and does not discuss DeepX.
The key insight: As we read it, DeepX is spending ahead of its sales: more than $100 million on the DX-M2 against about $13.3 million of orders to October and a $40 million revenue forecast for this year, on a chip meant to run models of about 100 billion parameters on about 5 W.
DeepX has finished the design of the DX-M2 and sent the final plans to Samsung’s foundry unit for production, Bloomberg reported, citing a statement from DeepX. It said the company expects its first silicon by March 2027, before customer evaluations and mass manufacturing in 2028.
The cost of more than $100 million covers performance verification and production engineering expenses, according to Quartz’s summary of the Bloomberg report. Quartz says DeepX is aiming the chip at robotics, factory and other industrial machinery, and PCs with built-in AI capabilities.
TheElec reported that the chip uses Samsung’s second-generation 2nm process, SF2P, and that DeepX completed the mask tape-out for prototype production in early November. It reported evaluation samples for customers in the second quarter of 2027, initial mass-production shipments in the fourth quarter and full-scale production from 2028. Our piece on Samsung’s US 2nm chip deal covered another 2nm customer for Samsung.
DeepX’s DX-M2 against its first chip, the DX-M1, with each figure shown as a multiple of the DX-M1, from DeepX’s product pages; the bandwidth figures are from DeepX’s CEO in an EE Times interview. The DX-M2 figures are DeepX’s targets and subject to change; the multiples are our arithmetic.
The DX-M2 Against the DX-M1
The DX-M1, made on Samsung’s 5nm process, went into mass production in August 2025, TheElec reported. Its product page lists 25 TOPS of INT8 compute, up to 8GB of memory and language models of up to 4B parameters, on an M.2 module rated at 5W maximum.
The DX-M2 page lists 80 TOPS, 153.6 GB/s of memory bandwidth, about 30 tokens per second at single batch, and up to 96GB of memory on the DX-M2 FcBGA package, 12 times the DX-M1’s 8GB, by our arithmetic. The page says the specifications “are subject to change without notice during development”.
DeepX’s founder and CEO, Lokwon Kim, told EE Times: “The most important change [between M1 and M2] is the memory bandwidth”. EE Times reported that bandwidth tripled from 50 to 150 GB/s and that the chip runs models of 20 billion to 100 billion parameters at 10 to 30 tokens per second, within the same 5 W budget as the M1.
The Company Behind the Bet
TheElec reported that DeepX forecasts revenue of $40 million this year, more than double the $17 million target it set last year, after revenue of 3.3 billion won and an operating loss of 20.1 billion won in 2025. More than $100 million is 2.5 times that forecast, by our arithmetic.
Kim told EE Times that DeepX has more than 100 purchase orders for the DX-M1. The company’s 10 November statement, as reported by TheElec and Chosunbiz, gave 88 orders worth about $13.3 million as of October. The two counts come from different dates and sources.
Kim framed the bet as a way around data center limits. “If we deploy models up to 100 billion parameters in devices, we can reduce AI data center traffic by more than 80%”, Kim told EE Times. That is the company’s claim, not a measured result. Our piece on Samsung’s Texas fab tracks the foundry side.
The Structural Read
As we read the spec pages, the jump is in memory more than in compute: the model size DeepX targets rises 25 times and the maximum memory 12 times, while NPU compute rises 3.2 times, by our arithmetic.
As we read the timeline, the money is committed before the first silicon: Bloomberg reported first silicon by March 2027, and the reports we read place volume production between late 2027 and 2028.
Lokwon Kim, DeepX founder and CEO, to EE Times
“This is important because with 5 W power consumption we can go everywhere, drones or any other battery-powered device”
Three Implications
FOR BUYERS OF EDGE AI CHIPS DeepX lists the DX-M2 at 80 TOPS and about 5 W, but its page says the specifications are subject to change during development, and first silicon is expected in March 2027.
FOR SAMSUNG’S FOUNDRY TheElec reported the DX-M2 uses Samsung’s SF2P process; our piece on Samsung’s Texas fab tracks the foundry’s US build.
FOR DEEPX’S FINANCES More than $100 million is 2.5 times the $40 million revenue DeepX forecasts for this year, by our arithmetic, after an operating loss of 20.1 billion won in 2025, TheElec reported.
The framework records: “AI has crossed from the training era to the inference era. This is not a technical nuance.”
As we read it, the DX-M2 is a bet on where that inference runs: on a 5 W chip inside the device rather than in a data center. The framework does not discuss DeepX.
DeepX’s planned spend on the DX-M2 as reported by Bloomberg, against its 2026 revenue forecast and its purchase orders to October from its statement as reported by TheElec. The multiples are our arithmetic.
What Is Not Established
We could read only the opening of Bloomberg’s article; the cost breakdown comes from Quartz’s summary of it. The parts we read do not say how DeepX will pay for the chip. The DX-M2 figures are DeepX’s own targets for a chip that has not yet produced first silicon.
The reports differ on timing: Bloomberg gives mass manufacturing in 2028, EE Times gives mass production targeted for fall 2027, and TheElec gives initial shipments in the fourth quarter of 2027. DeepX was not contacted for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
DeepX plans to spend more than $100 million on a 2nm chip made by Samsung, 7.5 times the about $13.3 million of orders it had secured by October and 2.5 times its $40 million revenue forecast for this year, by our arithmetic. On its own figures the DX-M2 offers 3.2 times the DX-M1’s compute at about the same 5 W.
A note on sourcing. Bloomberg reported the plan on 13 November; we could read only the opening of its article, and the cost breakdown and target markets come from Quartz’s summary of that report. The chip specifications are from DeepX’s own DX-M2 and DX-M1 product pages, read on 14 November, and from DeepX’s CEO in an interview with EE Times. The orders, revenue forecast and timeline are from DeepX’s 10 November statement as reported by TheElec and Chosunbiz. The multiples are our arithmetic. DeepX was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
