Groq calls itself the premier neocloud for fast inference and describes one platform for infrastructure, inference, and control. On the platform page that stack is named: GroqMetal provides infrastructure, GroqCore adds inference, and GroqAssured adds enterprise controls. Each includes the layers below. The same rack figures and says it is now operating 13 data centers across four continents.
Groq fits a buyer who is choosing how much of an inference stack to operate, not someone shopping a self-serve model menu. GroqMetal is the dedicated bare-metal layer. GroqCore is the tested stack for people who do not want to run that metal. GroqAssured adds governance and auditability. It is a weak fit if you need a published price, a model list, or the text of the privacy policy before you talk to the company. Those items are not in the public pages. LPX is described as a rack that connects LPU accelerators to NVIDIA Vera Rubin when released, so it should not be read as a rack the page says is already online.
A good fit for
Teams that want dedicated bare-metal inference capacity and the control that GroqMetal describes.
Teams that want production inference without operating the infrastructure, which is how GroqCore is worded.
Organizations that want governance, auditability, and control named as a separate layer.
Know the trade-offs
.
The other two are not named here.
LPX is introduced with the words when released. Do not treat that rack as current capacity.
Metal, Core, Assured, and the published rack.
Three layers, each including the one below
The platform page tells you to build a stack rather than pick a single feature. GroqMetal provides infrastructure, GroqCore adds inference, and GroqAssured adds enterprise controls. Each includes the layers below. That sentence is the only map of what sits inside what. Metal is the bottom layer, Core adds inference on top of it, and Assured adds controls on top of both. GroqMetal is dedicated bare-metal infrastructure, tuned for speed and reliability, with full control in your hands. GroqCore is a tested inference stack that turns dedicated capacity into production-ready performance, no infrastructure expertise required. GroqAssured is enterprise-grade governance, auditability, and control layered on top, so scale never comes at the cost of trust.
Under the layer names, the platform page repeats one set of rack figures: 256 LPUs per rack, 40 PB/s SRAM bandwidth, 1,000 tokens/sec/user, 128 GB of on-chip SRAM per rack, and 315 PFLOPS of FP8 inference compute. The same five lines are shown more than once. This profile treats them as the page's published numbers. The homepage says Groq pioneered the LPU and that inference, not training, is where the page locates the value. It also says fast or affordable is no longer a tradeoff, and that the company is building hundreds of megawatts of capacity. None of those sentences is tied to a dollar rate, a model name, or a per-token meter. The token-per-second figure is per user, as shown, and the page does not say how many users share a rack.
Named sites, and an LPX rack marked as not yet released
The platform page says it is now operating 13 data centers across four continents. The names are US-1 Liberty Lake, US-2 St Paul, US-3 Dallas, and US-4 Houston in the United States; CA-1 Kamloops, CA-2 Vaudreuil-Dorion, and CA-3 Calgary in Canada; EU-1 Vantaa in Finland and EU-2 London in the UK; AU-1 Sydney in Australia; and SA-1 Riyadh in Saudi Arabia. That is 11 labeled sites. A separate block describes LPX. When released, each NVIDIA Groq 3 LPX rack connects 256 LPU accelerators to NVIDIA's Vera Rubin to deliver low-latency, large-context inference for production AI. The page also repeats the same rack figures beside that block. Because the sentence starts with when released. The homepage mentions LPX working alongside NVIDIA GPUs. It does not give a date.
Those are the site's own scale lines. The body defines inference with a list of outcomes, from a customer served to an agent task completed, and says training creates the possibility while inference creates the value. It does not show a console, a request format, or a sample response. The legal page is an index. It links Terms of Use, Privacy Policy, Cookie Policy, Security, Trademark Policy, Photography and Filming Policy, and Recruitment Fraud Awareness. The public text stops at the titles. It does not include retention rules, a security control list, or the terms that would govern a GroqCore deployment. Start building is a link label on both pages. Was not verified here.
A team that already runs its own inference jobs, and wants dedicated bare metal with full control, matches the wording of that layer. The published rack is the only capacity sketch on the page: 256 LPUs, the SRAM figures, and the FP8 compute number. The page does not say how you reserve a rack, which of the named cities you can choose, or what full control includes. Treat the city list as locations the company says are operating, not as a menu you have already been shown.
Hand the infrastructure to GroqCore
For a group without infrastructure staff is GroqCore. The page calls it a tested inference stack that turns dedicated capacity into production-ready performance, and it says no infrastructure expertise is required. That is the contrast with Metal: capacity is still dedicated, but the stack is operated for you. Nothing names the models on that stack, the request shape, or a rate limit. If those details decide the purchase, they are not on the pages this profile uses. The only performance line attached to the hardware is 1,000 tokens per second per user.
Add the governance layer
A suggested reason to look at GroqAssured is the sentence on governance, auditability, and control. The page places it on top of the other layers, so it is not a separate cloud. A buyer who needs an audit story would start there and then ask what the unnamed controls actually are. The legal index shows that Security and Privacy Policy exist as documents. It does not quote them. Do not fill that gap with assumptions about certifications or retention. The public pages simply do not say.
How to get started with Groq
1
Pick the layer
Read GroqMetal, GroqCore, and GroqAssured as a stack. Each higher layer includes the layers below.
2
Read the rack lines
Note 256 LPUs per rack, the SRAM bandwidth and capacity, 1,000 tokens/sec/user, and 315 PFLOPS of FP8 compute.
3
Check the site list
Compare the 13-center claim with the 11 cities named, and keep LPX in the when-released column.
A price was not verified here. Ask the vendor before you pay. A free tier, a paid tier, and a token rate were not verified on the homepage, platform page, or legal index reviewed here.
The homepage says inference should be affordable and that fast or affordable is no longer a tradeoff. That is not a rate card. A funding headline is also on the pages. GroqAssured is described as enterprise controls without a quoted enterprise fee.
GroqMetal provides infrastructure, GroqCore adds inference, and GroqAssured adds enterprise controls. Each includes the layers below. Metal is dedicated bare metal with full control. Core is the tested stack that does not require infrastructure expertise. Assured adds governance, auditability, and control.
It includes the platform 256 LPUs per rack, 40 PB/s SRAM bandwidth, 1,000 tokens/sec/user, 128 GB of on-chip SRAM per rack, and 315 PFLOPS of FP8 inference compute. The page does not attach them to a named model.
It says Groq is now operating 13 data centers across four continents, then lists Liberty Lake, St Paul, Dallas, Houston, Kamloops, Vaudreuil-Dorion, Calgary, Vantaa, London, Sydney, and Riyadh. The public list is shorter than 13. The homepage adds that hundreds of megawatts are being built.
No. The platform sentence begins with when released. It says each NVIDIA Groq 3 LPX rack connects 256 LPU accelerators to NVIDIA's Vera Rubin to deliver low-latency, large-context inference for production AI. The homepage also mentions LPX beside NVIDIA GPUs, without a ship date.
Do the public pages include a price or the privacy policy?
No price is published on the public pages. The legal index lists Privacy Policy, Terms of Use, Cookie Policy, Security, Trademark Policy, Photography and Filming Policy, and Recruitment Fraud Awareness, and does not include their text. Model names, rate limits, and signup fields are also absent.
NavHunt did not run Groq. This profile restates the homepage, platform page, and legal index public from groq.com on 6 October 2026. It is not a test of token speed, rack capacity, or the Start building flow. This description is based on Groq's official materials reviewed on 2026-10-06, including Groq homepage, Groq platform, Groq legal index.