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

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 ^[7], and Nvidia packages 256 of them into an LPX rack ^[6], 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 ^[11]. Senior director Dion Harris put it as "using the right price, right processor for the right part of the workload" rather than replacing GPUs ^[10]. 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" ^[15].

The 3,400 tokens per second Nvidia cites is a full-rack number from an Artificial Analysis benchmark ^[6]. OpenAI's Ultrafast mode is quoted at 750 tokens per second on Cerebras silicon ^[13]. 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 ^[12].

The quieter point is the fab. Samsung builds Groq's chips while TSMC builds Nvidia's GPUs ^[8]. Wedbush's Matt Bryson, who rates the stock Outperform, argues that "component and material access, not end demand, is defining shipments" ^[17], 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 ^[16][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 ^[9], is still the vendor's own assertion about what buyers will pay.