---
title: "Supermicro will buy 10,000 SambaNova chips to write AI answers alongside Nvidia; gross margin fell to 12.7%"
description: "In the same week, the two Nvidia box sellers gave SambaNova and Groq the same job: the decode step in racks that carry Nvidia silicon and the OEM's own badge, with no published price, spec or end customer. Supermicro committed 10,000 chips against a quarter of 12.7 percent margins and $434 million of cash burn."
publisher: "The Inference"
section: "Chips"
published: 2026-11-06T15:19:31.915Z
modified: 2026-11-06T15:19:31.915Z
canonical: https://theinference.org/article/supermicro-will-buy-10-000-sambanova-chips-to-write-ai-answers-alongside-nvidia-gross-margin-fell-to-12-7
language: en
keywords: "AI Chips, Semiconductors, Inference, Earnings, Data Centers"
---

# Supermicro will buy 10,000 SambaNova chips to write AI answers alongside Nvidia; gross margin fell to 12.7%

> In the same week, the two Nvidia box sellers gave SambaNova and Groq the same job: the decode step in racks that carry Nvidia silicon and the OEM's own badge, with no published price, spec or end customer. Supermicro committed 10,000 chips against a quarter of 12.7 percent margins and $434 million of cash burn.

[Supermicro](https://theinference.org/markets/companies/super-micro), the San Jose server maker that builds most of [Nvidia](https://theinference.org/markets/companies/nvidia)'s AI boxes, [said this week it will buy 10,000 SN50 chips](https://www.datacenterdynamics.com/en/news/sambanova-partners-with-supermicro-for-sn50-rdu-deployments/) from [SambaNova](https://theinference.org/markets/companies/sambanova), the San Jose AI chip company founded in 2017, and build inference racks that split the work: Nvidia GPUs read the prompt in the prefill step, SambaNova's Reconfigurable Dataflow Units write the answer in the decode step [1][2]. The racks carry Supermicro's badge inside its Data Center Building Block Solutions line, and Supermicro handles integration, testing, delivery and on-site service [1][3].

[Days earlier at GTC Berlin](https://www.storagesserver.com/cases/dell-ai-factory-with-nvidia-adds-a-256-cpu-vera-rack-a-groq-3-lpu-inference-rack-and-a-64gb-gb10-des-102861.html), Dell, the other big Nvidia box seller, added its first [Groq](https://theinference.org/markets/companies/groq) rack: the PowerEdge XE9810, sixteen direct-liquid-cooled trays of Groq 3 LPUs, shipping in the first quarter of 2027 [6].

## The order has no price, no spec and no customer

Neither company published what one SN50 costs, what it delivers in tokens per second, or who buys the first rack [1]. SambaNova's own benchmark, 5x faster than "competitive chips" at 3x lower total cost of ownership, names no comparator [1]. The markets targeted first, neoclouds and sovereign clouds, have not ordered publicly [2].

What landed instead is the buyer's balance sheet. Supermicro's [quarter ended September 30 showed net sales of $19.946 billion](https://www.stocktitan.net/sec-filings/SMCI/10-q-super-micro-computer-inc-quarterly-earnings-report-458aa59c4835.html), four times the year-ago $5.018 billion, with gross profit of $2.541 billion, a 12.7 percent margin against 17.5 percent the quarter before [4][5]. Inventories rose to $15.054 billion from $12.896 billion in three months [4]. Operating activities used $434.0 million [4]. One customer was 44.0 percent of quarterly sales [4]. In June the company sold stock for about $1.4 billion combined [4].

*Chart: **Supermicro's sales quadrupled in a year** Line chart of Supermicro quarterly net sales in billions of dollars across five quarters, rising from 5.0 in the quarter ended September 2025 to 19.9 in the quarter ended September 2026.*

*Quarterly net sales in billions of dollars for the fiscal quarters ended September 2025 through September 2026. Sources: Super Micro Form 10-Q filed 5 November 2026; Super Micro, quarterly figures from its SEC filings. [4][9]*

*Chart: **Gross margin swung from 6.3% to 17.5% and back to 12.7% in five quarters** Line chart of Supermicro gross margin in percent across five fiscal quarters, swinging from 9.3 percent down to 6.3, up to 9.9, up to 17.5 and back to 12.7.*

*Gross margin as a percent of net sales for the same fiscal quarters. The 12.7 percent quarter is the one in which the SambaNova order landed. Sources: Super Micro Form 10-Q filed 5 November 2026; Super Micro, quarterly figures from its SEC filings. [4][9]*

The 10,000 RDUs are inventory bought ahead of demand the sources do not show. That is the deal's real shape.

## Decode has other bidders

The decode slot SambaNova just took is contested. [Qualcomm](https://theinference.org/markets/companies/qualcomm) [said on its November 4 earnings call](https://finance.yahoo.com/markets/stocks/articles/qualcomm-inc-qcom-q4-2026-051630325.html) that decode is the workload drawing the strongest customer feedback for its custom data-center silicon, and guided its first data-center quarter past $650 million [7].

[Nvidia itself paid $13 billion for Groq](https://theinference.org/article/cerebras-s-customer-pays-3-to-5x-for-2-000-tokens-a-second-culper-research-says-the-wafers-fail-in-2-years) to add a fast-inference accelerator [8]. A paying customer told The Inference that a fast token sells for three to five times a standard one [8], so the slot is worth money to whoever holds it. The challengers are not being embraced; they are being rented.

## Three tells decide whether the order is real

First deployments go to neoclouds and sovereign clouds, with enterprise sales to follow [1]. The tells: a named end customer, a published throughput figure for SN50, and whether Supermicro's gross margin holds near 12.7 percent while it carries the inventory [1]. None of the three exists today.

## Takeaway

Supermicro committed 10,000 SambaNova chips for the decode step with no published price, spec or end customer, in the same quarter its gross margin fell to 12.7 percent and operations burned $434 million.

## Sources

1. [Data Center Dynamics, SambaNova partners with Supermicro for SN50 RDU deployments, 4 November 2026](https://www.datacenterdynamics.com/en/news/sambanova-partners-with-supermicro-for-sn50-rdu-deployments/)
2. [Data Centre Magazine, Supermicro's New Inference Racks to Use SambaNova SN50 Chips, collected 5 November 2026](https://datacentremagazine.com/news/supermicros-new-inference-racks-to-use-sambanova-sn50-chips)
3. [MarketScreener, Super Micro Computer enters strategic collaboration with SambaNova Systems, 6 November 2026](https://www.marketscreener.com/news/super-micro-computer-inc-enters-strategic-collaboration-with-sambanova-systems-inc-to-develop-sn-ce785fdfd18df125)
4. [Super Micro Computer Form 10-Q, quarter ended 30 September 2026, via StockTitan, filed 5 November 2026](https://www.stocktitan.net/sec-filings/SMCI/10-q-super-micro-computer-inc-quarterly-earnings-report-458aa59c4835.html)
5. [InsiderFinance, Supermicro Earnings Rise on AI Demand, 3 November 2026](https://www.insiderfinance.io/news/supermicro-earnings-rise-on-ai-demand)
6. [StorageSServer, Dell AI Factory with NVIDIA Adds a 256-CPU Vera Rack, a Groq 3 LPU Inference Rack, and a 64GB GB10 Desktop, 3 November 2026](https://www.storagesserver.com/cases/dell-ai-factory-with-nvidia-adds-a-256-cpu-vera-rack-a-groq-3-lpu-inference-rack-and-a-64gb-gb10-des-102861.html)
7. [GuruFocus via Yahoo Finance, Qualcomm (QCOM) Q4 2026 Earnings Call Highlights, 5 November 2026](https://finance.yahoo.com/markets/stocks/articles/qualcomm-inc-qcom-q4-2026-051630325.html)
8. [The Inference, Cerebras's customer pays 3 to 5x for 2,000 tokens a second, 4 November 2026](https://theinference.org/article/cerebras-s-customer-pays-3-to-5x-for-2-000-tokens-a-second-culper-research-says-the-wafers-fail-in-2-years)
9. [Super Micro, quarterly figures from its SEC filings, retrieved 6 November 2026](https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0001375365)


---

Supermicro will buy 10,000 SambaNova chips to write AI answers alongside Nvidia; gross margin fell to 12.7% — The Inference. Canonical: https://theinference.org/article/supermicro-will-buy-10-000-sambanova-chips-to-write-ai-answers-alongside-nvidia-gross-margin-fell-to-12-7