DeepX will spend more than $100 million developing a next-generation AI chip fabricated on Samsung Electronics' 2-nanometer process, betting that robots and factory machines will need powerful intelligence baked directly into the hardware rather than fetched from distant data centers.

The South Korean startup has finished the design work. It handed the completed blueprints for the chip, called the DX-M2, to Samsung's contract-manufacturing arm — the foundry business that builds chips for other companies — so production can begin, Bloomberg reports.

The timeline is tight and specific. DeepX expects its first working silicon by March 2027. Customer testing follows, and large-scale production is planned for 2028.

What does the money actually buy?

The $100 million-plus figure covers the unglamorous middle stretch of chip development, not just design. According to Bloomberg, the money goes to performance verification (proving the chip actually works as specified) and production engineering (adapting the design so a factory can manufacture it reliably at volume).

That's the expensive part. A design on paper costs relatively little. Turning it into something Samsung's 2-nanometer fabrication lines (currently the company's most advanced chipmaking node) can stamp out by the thousands is where budgets balloon.

Why on-device AI instead of the cloud?

The strategic logic behind the DX-M2 is a wager on where AI computing happens. DeepX believes AI models run better on the devices themselves than on servers in far-off data centers, per Bloomberg.

The reasoning comes down to three measurable costs of cloud computing:

• Power: every round trip to a data center consumes electricity on both ends
• Expense: cloud inference (running an AI model on remote hardware) means paying for someone else's servers, continuously
• Latency: data traveling to the cloud and back introduces delays — a serious problem for machines that must react in real time

For a factory robot or an industrial controller, a 200-millisecond cloud round trip can be the difference between a smooth motion and a crash. On-device processing removes that variable entirely.

There's also a political dimension. Data centers are drawing growing pushback over how much electricity and water they consume, Bloomberg notes. DeepX's semiconductors deliberately target AI tasks outside those facilities, positioning the company on the right side of an intensifying resource debate.

How efficient is 5 watts?

The DX-M2's headline engineering claim: it handles generative-AI workloads — the same class of models behind chatbots and image generators — while drawing only about 5 watts of power.

That number matters because it's a fraction of what a data-center GPU consumes. A chip sipping 5 watts can run inside a robot, a piece of factory machinery, or a laptop-class device without exotic cooling or massive batteries. It's the difference between AI as a bolted-on cloud service and AI as an embedded component, silent and always available.

Fitting that capability onto Samsung's 2-nanometer process is what makes the target plausible. Smaller transistors mean more compute per watt, and the 2nm node represents the current frontier of that shrink.

Which markets is DeepX chasing?

The startup is aiming the DX-M2 at three segments, per Bloomberg:

• Robotics — machines that need onboard intelligence for perception and control
• Factory and industrial machinery — equipment where downtime and latency carry hard costs
• AI-capable PCs — personal computers with built-in AI acceleration

That's a broad spread for a single chip, but the common thread is clear: all three need serious AI compute in a tight power envelope, and none can afford to depend on a network connection.

What happens between now and 2028?

The critical window runs from March 2027, when first silicon arrives, through customer testing, to volume production in 2028. Tape-out — the industry term for handing final designs to a foundry — is done. What remains is execution risk: will the chip hit its 5-watt target on real workloads, and will customers sign on during the testing phase?

By 2028, DeepX will know whether its $100 million bet on edge AI (computing done on the device rather than in the cloud) was early or exactly on time. Given how fast device makers are pushing intelligence into hardware, the timing looks, at minimum, well chosen.