Invest Independently confirmed 2 publishers3 min readPublished Aug 24, 2026, 10:41 AM PDT

Nvidia's $20bn Groq buy becomes shipping racks, and the tape reads it as execution risk

Hardware from the largest acquisition in Nvidia's history is in production and headed to a neocloud this year. The market sold it anyway, two days before earnings.

The Investor · Invest desk

How we use AI Send a correction

What happened

Nvidia said Monday that its Groq 3 LPX rack is in full production, commercialising the technology behind its largest acquisition on record.
The racks go to neocloud Nebius alongside Vera CPUs and Rubin GPUs, and will be online later this year, according to senior director Dion Harris.
The stock opened about 3% higher and finished roughly 2% lower after the Groq announcement.
Nvidia reports fiscal second-quarter results on Wednesday after the market closes.

Why it matters

capability Neoclouds acquire a product line they could not price before: a latency-tiered service sitting above standard token serving, which turns response time into a billable feature rather than an...
constraint Huang has already capped the Groq line himself at a quarter of coding-workload data centre space, so the return on the purchase has to come from protecting GPU pricing at the high end, not from...
exposure Part of Nvidia's inference roadmap now sits at Samsung rather than TSMC, which is useful capacity if components are the binding constraint and a new dependency if they are not.
contradiction Two published estimates of the same share price are a quarter apart, and Monday's trading agreed with neither, which is the definition of a name where the debate is execution rather than demand.

Roughly five points of intraday reversal on a company valued near $5.2 trillion is about $260 billion of market capitalisation moving on one product announcement ^[21], and it came in a month when the shares had been running nearly 7% against 2.5% for the S&P 500 ^[20].

What the buyer of that stock is underwriting here is smaller than the headline suggests. Nvidia paid $20 billion for Groq's assets in December ^[3]. Set against the $1 trillion in cumulative Blackwell and Vera Rubin sales Jensen Huang projected through 2027 when the products were unveiled in March, the deal is about 2% of the revenue Nvidia has already pointed the market at ^[23]. Groq does not need to be a large business. It needs to stop someone else from owning the fast end of serving.

The hardware is narrower than a GPU by design. Each Groq 3 chip carries 500 megabytes of SRAM on the die itself to cut memory bottlenecks ^[6], and Nvidia packages 256 of them into an LPX rack ^[5], so a full rack holds roughly 128 gigabytes of on-chip memory ^[22]. That is a decode-phase machine, which is exactly what Nvidia says it is: low-latency parts handle the decode portion of serving, while GPUs still do training and adapt to new models ^[10]. Senior director Dion Harris put it as "using the right price, right processor for the right part of the workload" rather than replacing GPUs ^[9]. Huang's own allocation was more specific about the ceiling: a quarter of the data centre space intended for coding goes to Groq, and "the rest of my data center is all 100% Vera Rubin" ^[12].

The 3,400 tokens per second Nvidia cites is a full-rack number from an Artificial Analysis benchmark ^[5]. OpenAI's Ultrafast mode is quoted at 750 tokens per second on Cerebras silicon ^[16]. The division gives 4.5x ^[26], but the sources do not say the two figures are measured on the same basis, and one is a rack while the other is a service tier, so it is a marketing comparison rather than a competitive one. AMD has already said it will pair its rack-scale systems with Cerebras chips for the same low-latency work ^[15].

The quieter point is the fab. Samsung builds Groq's chips while TSMC builds Nvidia's GPUs ^[7]. Wedbush's Matt Bryson, who rates the stock Outperform, argues that "component and material access, not end demand, is defining shipments" ^[18], which makes a Samsung-fabbed product line a second source of units as much as a latency play.

The sell side shows how wide the range still is. Rosenblatt keeps a Buy and a $325 target against a $214.72 quote, about 51% above the tape ^[17]^[24], while InvestingPro's fair value of $259.96 sits 21% above the price, with Rosenblatt's target a quarter above that ^[25]. All of it rests on a consensus that already has quarterly revenue and earnings roughly doubling from a year ago ^[19]. And the premium-tier logic, that cloud providers can charge more for tokens carrying stricter latency terms ^[8], is still the vendor's own assertion about what buyers will pay.

What to watch

Whether the Nebius deployment actually goes live this calendar year, and whether a second neocloud customer is named for LPX racks.
Wednesday's October-quarter guide: above the Street as Wedbush expects, or merely in line with it.
Whether Nvidia ever breaks out Groq-line revenue rather than folding it into the data centre segment, where its scale cannot be checked.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence58
Adoption28
Hype gap+24
Incentives74
Confidence61

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Nvidia announced on Monday that its Groq 3 LPX rack is in full production, marking the commercialization of technology from the company's largest acquisition on record.
ReportedSupported2 sources— create a free account to open them View cited source
[2]
The Groq rack will be deployed alongside Vera central processors and Rubin graphics processors at neocloud Nebius and will be online later this year, Nvidia senior director Dion Harris told reporters.
ReportedSupported2 sources— create a free account to open them View cited source
[3]
In December, Nvidia bought assets from chip startup Groq for $20 billion, the company's largest purchase.
ReportedSupported2 sources— create a free account to open them View cited source
[4]
Nvidia fell about 2% Monday after opening roughly 3% higher; the stock jumped about 3% soon after the opening bell, then dropped roughly 2% after the Groq news came out.
ReportedSupported2 sources— create a free account to open them View cited source
[5]
Nvidia packages 256 individual Groq 3 chips into its LPX racks and says the rack can deliver 3,400 tokens per second, citing a benchmark from Artificial Analysis.
ReportedSupported2 sources— create a free account to open them View cited source
[6]
The Groq architecture includes 500 megabytes of fast SRAM on the chip's die itself to reduce memory-related bottlenecks.
ReportedSupported2 sources— create a free account to open them View cited source
[7]
Groq chips are manufactured by Samsung, while Taiwan Semiconductor Manufacturing makes Nvidia's GPUs.
ReportedSupported2 sources— create a free account to open them View cited source
[8]
Cloud companies can charge more for low-latency tokens, Nvidia says; Harris said it "unlocks the ability to offer premium tiers of service for those users and those customers who actually demand the most latency-sensitive" service agreements.
ReportedSupportedSource: Nvidia senior director Dion Harris2 sources— create a free account to open them View cited source
[9]
"This isn't about replacing GPUs," Harris said. "It's about using the right price, right processor for the right part of the workload."
ReportedSupportedSource: Nvidia senior director Dion Harris2 sources— create a free account to open them View cited source
[10]
Low-latency chips do not replace GPUs, which can do training as well as inference and are flexible enough to adapt to new technologies and models; low-latency chips like Groq mainly focus on the "decode" phase of serving models.
ReportedSupported2 sources— create a free account to open them View cited source
[11]
At the Vera Rubin and Groq 3 LPX unveiling in March, Jensen Huang projected $1 trillion in cumulative sales between current-generation Blackwell chips and the new Vera Rubin systems through 2027.
ReportedSupported2 sources— create a free account to open them View cited source
[12]
Huang said he would allocate a quarter of data center space intended for coding applications to Groq chips: "The rest of my data center is all 100% Vera Rubin."
ReportedSupported2 sources— create a free account to open them View cited source
[13]
Nvidia is due to report its fiscal second-quarter numbers on Wednesday after the market closes.
ReportedSupported2 sources— create a free account to open them View cited source
[14]
Nvidia is worth close to $5.2 trillion based on FactSet data, making it the biggest company worldwide by market value.
ReportedSupportedView cited source
[15]
AMD announced earlier this year it would integrate its rack-scale systems with chips from Cerebras, which recently went public, focusing on low-latency inference.
ReportedSupportedView cited source
[16]
OpenAI's newly announced Ultrafast mode currently promises 750 tokens per second and is "powered by Cerebras."
ReportedSupportedView cited source
[17]
Rosenblatt Securities kept its Buy rating on Nvidia with a $325.00 share-price target; Nvidia was trading at $214.72, while InvestingPro put its fair-value estimate at $259.96.
ReportedSupportedView cited source
[18]
Wedbush analyst Matt Bryson, who has an Outperform rating, expects Nvidia to exceed its guidance and guide October above the Street, citing "a supply position we continue to view as the best in the industry at a point where component and material access, not end demand, is defining shipments."
ReportedSupportedSource: Wedbush analyst Matt BrysonView cited source
[19]
London Stock Exchange Group data shows analysts expect Nvidia's quarterly revenue and earnings to reach about twice their levels from the same period last year.
ReportedSupportedView cited source
[20]
Nvidia shares were up nearly 7% in August through Friday's close, while the S&P 500 had risen around 2.5%.
ReportedSupportedView cited source
[21]
A reversal from roughly plus 3% to roughly minus 2% is about five percentage points, or approximately $260 billion of market value on a roughly $5.2 trillion capitalisation.
DerivedSupported1 source— create a free account to open them
[22]
An LPX rack of 256 Groq 3 chips carries roughly 128 gigabytes of on-die SRAM in total.
DerivedSupported1 source— create a free account to open them
[23]
The $20 billion Groq purchase equals about 2% of the $1 trillion cumulative Blackwell and Vera Rubin sales figure Huang projected through 2027.
DerivedSupported1 source— create a free account to open them
[24]
Rosenblatt's $325 target is about 51% above the $214.72 trading price.
DerivedSupported1 source— create a free account to open them
[25]
InvestingPro's $259.96 fair value is about 21% above the $214.72 price, and Rosenblatt's $325 target is about 25% above that fair value.
DerivedSupported1 source— create a free account to open them
[26]
Nvidia's cited 3,400 tokens per second per rack is 4.5 times the 750 tokens per second quoted for OpenAI's Ultrafast mode, though the sources do not state that the two figures share a measurement basis.
DerivedSupported1 source— create a free account to open them

Sources

2 independent publishers whose own reporting we read for this story.

cnbc.com
1 article · August 24, 2026
Nvidia says Groq racks will be online this year following $20 billion purchase
cryptopolitan.com
1 article · August 24, 2026
Why is Nvidia stock falling today?

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

For builders, the substance is workload segmentation. A Groq 3 LPX rack is 256 chips with 500MB of on-die SRAM each (roughly 128GB of SRAM per rack), and Nvidia's own framing limits it to the decode phase of serving models rather than training or general flexibility: 'This isn't about replacing GPUs.' The headline 3,400 tokens per second is vendor-cited from Artificial Analysis, and the sources never establish that it is measured the same way as the 750 tokens per second attributed to OpenAI's Cerebras-powered Ultrafast mode, so the 4.5x ratio is arithmetic, not a like-for-like result. The competitive shape is clear: AMD plus Cerebras on one side, Nvidia plus Groq on the other, both aimed at latency-sensitive agentic and coding workloads. Nothing in the material shows a builder running code on this hardware yet.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence58: 0100
Named sourcing, single-origin technical claims
Core facts are on the record: an Nvidia senior director briefing reporters, an attributed CEO projection from March, and financial data attributed to FactSet, LSEG and named brokers. But the performance number is vendor-cited rather than independently reproduced, the second publisher largely restates the first, and the causal link between the announcement and the share move rests on one outlet's sequencing.
Adoption28: 0100
Production announced, one named customer, nothing live
Adoption evidence is a production declaration plus a single named neocloud (Nebius) with a 'later this year' go-live. There is no disclosed capacity, customer count, revenue, price tier or usage figure, and the vendor's own capacity model caps Groq at a quarter of coding-oriented data-centre space. Rival Cerebras silicon is already behind a user-facing OpenAI mode, which is adoption for the category but not for this product.
Hype gap+24: −100 understated0 aligned+100 overstated
Milestone language runs ahead of deployed reality
'Full production' and a 4.5x-looking throughput headline sit against zero live capacity, one named customer, an unverified benchmark and no attributable revenue from a $20 billion purchase that equals roughly 2% of the vendor's own $1 trillion cumulative projection. The overstatement is modest rather than severe because the primary source itself supplies the limiting caveats — decode-phase only, not a GPU replacement, 25% of coding capacity — and the market's negative reaction ran counter to the promotional framing.
Incentives74: 0100
Vendor briefing plus long-side sell-side, ahead of earnings
Every substantive technical claim originates in an Nvidia-controlled press call two days before the company's own earnings print, and the performance figure is one the vendor selected. The financial layer is supplied by a Buy-rated broker with a $325 target, an Outperform-rated analyst forecasting a beat, and a subscription valuation tool, with the second publisher also promoting its own live earnings coverage and newsletter. These interests are disclosed in the sources but not interrogated by either outlet.
Confidence61: 0100
Facts firm, significance unresolved
The announcement, deployment target, acquisition price, specifications and market data are consistently reported across two sources and are unlikely to be wrong. What remains uncertain is what any of it is worth: no independent benchmark, no live deployment, no revenue attribution, and only one publisher's word that the Groq news caused the intraday reversal. Confidence in the ledger is therefore well above confidence in the story's implied significance.

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Low-Latency InferenceFollow
Inference Token PricingFollow
Nvidia Earnings And ValuationFollow
AI Chip M&AFollow
AI Accelerator HardwareFollow
Semiconductor Manufacturing And SupplyFollow