Supermicro to pair SambaNova SN50 chips with Nvidia GPUs, starting with 10,000

The 10,000 SN50 chips are an initial batch, with no disclosed order. Nvidia is building its own decode chip for the same role.

The 10,000 SN50 chips in SambaNova's announcement describe an initial batch, not a disclosed order. Supermicro will build and support paired systems, but no customer, purchase price or delivery date is given.

The systems keep Nvidia GPUs for prefill, which processes the prompt, and add SambaNova chips for decode, which generates the response. The question is whether specialized decode justifies its cost alongside the GPUs.

What was announced

SambaNova's 3 November announcement says Supermicro will integrate, validate, deliver and support systems pairing Nvidia GPUs and SambaNova SN50 chips. The phased rollout starts with an "initial batch" of 10,000 chips. SambaNova names NeoClouds, specialist GPU rental firms, and sovereign clouds as first target markets, with enterprises later.

No named customer, firm order value, delivery date or rack count is disclosed.

Supermicro chief executive Charles Liang says the arrangement "expands our broad DCBBS portfolio by enabling heterogeneous architectures". SambaNova's Rodrigo Liang says customers can order SN50 chips "alongside GPUs".

Supermicro adds validation, delivery and support to SambaNova's route to buyers. That could ease buying, but it does not show the batch will ship.

Why the GPUs stay

SambaNova separates inference into prefill and decode. Prefill processes the prompt and, in the company's account, suits GPU computing power. Decode produces output one token at a time and, SambaNova argues, is limited by memory movement.

Its SN50 is a reconfigurable dataflow unit aimed at that task. SambaNova says agent software spends most of its time in decode, which describes this system, not every workload.

The design must justify another chip type and its integration. Speed gives a buyer no advantage if utilization is poor or total cost per token is higher.

What has been measured

In June SambaNova demonstrated Nvidia B200 GPUs handling prefill with its earlier SN40 chip handling decode. It reported twice the speed of B200 GPUs alone, verified by Artificial Analysis according to the post, which names Together.ai as the first commercial customer. It was not a test of SN50 systems.

The SN50 setup followed on 8 July. A SambaNova account describes four Nvidia H200 GPUs for prefill and one rack of 16 SN50 chips for decode. It reports up to 850 tokens per second for short prompts and over 450 for long ones, with no baseline.

SambaNova also reported a SemiAnalysis benchmark of one SN50 rack on the same model. It showed about 800 tokens per second per user, or about 400 with more throughput. SambaNova calls this "early validation".

The June SN40 result has a GPU baseline; the SN50 results do not. All are speed measures on vendor chosen setups, not customer cost per token or proof that a Supermicro system beats a GPU only design.

SambaNova's September illustration targets a six month payback at an "illustrative" 70 percent billable utilization, with no hardware price, so it cannot be reproduced.

Racks, power and Oracle

SambaNova's product page specifies 16 SN50 chips in each air cooled rack. A 27 October release with Oracle puts average draw at 20 kW per rack. It says SambaNova and Oracle plan to deploy SN50 chips in Oracle Cloud Infrastructure, with no quantity or date.

On CTOL's arithmetic, if all 10,000 chips occupied racks of 16, that would be 625 racks. At the cited 20 kW each, their SN50 draw would be about 12.5 MW, excluding Nvidia GPUs. This is a sizing illustration, not a forecast. No such deployment is disclosed.

Supermicro's margins and inventory

Supermicro's preliminary update for the quarter ending 30 September estimates revenue of $18.5 billion to $19.5 billion, above guidance. It puts gross margin at 12.0 to 12.5 percent, against 17.5 percent in June. The figures are unaudited; full results were due after the close on 3 November.

The June financial results show inventory of $12.9 billion, up from $4.7 billion a year earlier. Cash was $7.5 billion against $8.7 billion of debt. Operating cash flow was negative $6.8 billion for the fiscal year, though the June quarter generated $747 million.

Supermicro also disclosed that its board was reviewing certain transactions "in connection with export-control issues". The available information does not link the review to SambaNova or a buyer.

With those margins and that inventory, another accelerator could broaden Supermicro's offering or tie up capital. The announcement gives no SN50 price, cost or payment terms, so the deal's revenue, profit and cash contribution cannot be estimated.

Nvidia is filling the same slot

Groq announced a nonexclusive technology licensing agreement with Nvidia on 24 December 2025, and its founder joined Nvidia. Groq disclosed no price, while Tom's Hardware puts the deal at $20 billion.

Nvidia's 16 March announcement described Groq 3 LPX racks handling decode alongside Vera Rubin GPUs, targeted for the second half of 2026. Nvidia claimed up to 35 times the throughput per megawatt, with no stated baseline. That cannot be compared with SambaNova's tokens per second.

AWS pairs Trainium hardware for prefill with Cerebras CS-3 systems for decode on Amazon Bedrock. AWS vice president David Brown promises inference "an order of magnitude faster". It uses its own prefill chip, not Nvidia GPUs.

AWS supports the architecture; Nvidia's Groq 3 LPX plan challenges SambaNova within it. The competition is over chips for decode; Nvidia GPUs remain in prefill.

What could happen next

The following scenarios are ranked by likelihood, without numerical probabilities.

Most likely, few trials turn into orders. If buyers still need cost evidence, adoption may stay at trials or small orders, which the missing customer and price disclosures favor. A named order could appear on the 3 November call or in the next filing.

An anchor order is also plausible. A NeoCloud or sovereign cloud could commit once cost per token meets its needs. By 30 June 2027, the sign would be a named buyer, quantity and delivery period. Oracle's plan does not qualify.

Broad adoption of the pairing as a standard design is unlikely. It requires repeat orders and a defensible cost comparison against Nvidia's Groq 3 LPX. By 30 June 2027 the sign would be named repeat customers and comparative economics, not another peak speed claim.

What would prove CTOL's prediction wrong?

CTOL expects with moderate confidence that by 30 June 2027 neither company will name a customer with a firm order for Supermicro SN50 systems. This concerns public disclosure, not total demand.

The basis is an announced batch with no named buyer, price or delivery date. A press release, filing or earnings call statement naming such a customer would prove the prediction wrong. An Oracle plan or a speed demonstration would not.

A firm order would prove a commitment, not superior economics. Rodrigo Liang says SambaNova looks forward to joining the official InferenceX benchmark, which could show speed but not a buyer's cost per token.