Skip to content
unnudged
Get early access
brandCooling

Groq on X

Builders credit Groq for what they shipped on it — the largest group, 21 people — with the free tier at 19. The recurring complaint is months of paid plans and billing switched off, and NVIDIA's licensing deal is still read as a takeover.

First written 27 Oct · what changed · from 126 posts by 113 people

Tracking posts that say “@GroqLLC” or “groq”

Track this for your market
Posts, last 24h
20
−49%vs. previous 24h
Posts, last 7 days
320
96 in the last 72h
Velocity
46/day
7-day average
Momentum
−56%
Last 24h vs. a normal day this week

Posts per day, last 30 days

as of Nov 3, 11:01 UTC
100500Oct 20: 41 postsOct 20Oct 21: 18 postsOct 22: 60 postsOct 23: 30 postsOct 24: 61 postsOct 25: 21 postsOct 26: 61 postsOct 26Oct 27: 68 postsOct 28: 51 postsOct 29: 72 postsOct 30: 42 postsOct 31: 34 postsNov 1: 41 postsNov 2: 30 postsNov 2

What changed

as of 27 Oct · checked once a month

First written 27 Oct. When what people say changes, it shows here.

Tone· 71 posts from 64 people· written 27 Oct

How do people feel about Groq?

Builders who use Groq recommend it — 21 people credit it inside apps, agents and pipelines they shipped, the biggest group in the reception talk — and the standing free tier is just behind at 19.

The biggest theme is that it is already working inside things people built: intent parsing in a Solana transfer bot, an AI research agent on Python and Tavily, a full-duplex voice pipeline running GPT-OSS-120B, Llama 3.3 scoring trading signals, a social-listening tool. The praise is for the job done rather than a benchmark.

The free tier is close behind: 19 people describe 200,000 tokens a day, 1,000 requests, 30 a minute, no credit card, and call it a "generous free tier" that carries a real workload rather than a demo. 12 people use the Whisper models for dictation and transcription, naming whisper-large-v3-turbo and citing free access and speed; one post claims an hour of audio transcribes in under three seconds. 10 praise raw speed, with one logged comparison putting Groq at 0.35s against 1.06s and 3.78s for other providers.

Not everyone is positive: 4 sound soured, one calling it "closed for business" and another citing an unreliable free tier beside an unavailable paid one.

  • Satisfied users recommend it23
  • Loves the free tier19
  • Great for Whisper transcription13
  • Praises the speed12
  • Disappointed with Groq4
See 5 of 71 posts
black panther@pantherbllack

1,200,000 FREE TOKENS A DAY ARE SITTING AT TWO PROVIDERS AND MOST PEOPLE KNOW NEITHER [where the 1.2M comes from] → Groq: 200,000 tokens a day, 1,000 requests, 30 a minute. a standing free tier, and they don't train on what you send link → Cerebras: 1,000,000 tokens a day, but out of a $5 grant that dies 30 days after it lands. needs a verified card, doesn't train link the first one lasts forever, the second one lasts a month. together that's 1.2M tokens a day until the clock runs out [4 more while you're there] → Cloudflare Workers AI: 10,000 neurons a day, 300 requests a minute link → OpenRouter: models ending in :free, 50 requests a day, or 1,000 once you've bought 10 credits link → Cohere: 1,000 calls a month link → Alibaba Model Studio: 1M tokens per model, 90 days, Singapore region only link [the trap nobody reads] Alibaba switches you to pay-as-you-go the moment the quota runs out, unless you turn on "Free Quota Only". it's off by default Google's unpaid quota (link trains on what you send. Mistral's free mode (link does too, by default, with an opt-out in the admin panel Groq, Cloudflare and Cerebras don't [how to wire it] → make Groq the default. it doesn't expire and doesn't train on your code → add Cerebras for heavy days while the grant is alive → keep OpenRouter as the fallback when you hit a wall → never send client code to a tier that trains on it you're not saving money on AI. you just never collected what's already free ↓

Paige@Paigedays911

Most developers have no idea this secret free tier exists. 👀🔥 A hardware startup called 𝗚𝗿𝗼𝗾 engineered dedicated AI chips (LPUs) that make standard cloud LLMs feel like dial-up internet. • Generates text 10x faster than human reading speed (800+ tokens/sec) • Permanent generous free tier giving 500,000+ tokens daily with $0 spend • Transcribes 1-hour audio files in under 3 seconds using Whisper Large • Works seamlessly with standard OpenAI Python and Node.js libraries Are you still paying OpenAI for basic agent loops, or have you switched to free LPUs yet? Console & setup: link

Vicarioy | ML Reasearcher@vicarioy_

Turns out it wasn’t random. OpenRouter rotates requests across upstream providers. I logged every call’s actual serving provider and ran 12 identical back-to-back requests: Groq averaged 0.35s. Novita averaged 1.06s. DeepInfra averaged 3.78s, for the exact same model, same prompt

Alok Ranjan@agentspanel

@typesafeai Jev powering near real time full duplex voice agent with cascaded architecture. Here's the pipeline: mic → @DeepgramAI Flux (turn detection) → @typesafeai Jev → @GroqLLC GPT-OSS-120B LLM → @SarvamAI TTS → speaker, with a live @excalidraw canvas on the side.

5 of the 71 posts behind this answer, one per person. The other 66, and every new one as it’s posted, come with tracking. Get early access

Pain points· 25 posts from 21 people· written 27 Oct

What do people complain about with Groq?

The largest complaint is that paid developer plans and billing are switched off — 4 people — level with the 4 reporting an outage, missing capacity or enabled models that never appear.

People asking about plans describe the developer tier as "switched off", pay-as-you-go pricing redirecting to the homepage, and waits of months; one asks to be allowed onto pay-as-you-go at all. Outages are the other half of the same size: a model down for days until a user moved a project to Claude, "No compute" on a given day, a broken link, models enabled in the console that never show up.

Behind those, 3 people cannot get in — a GitHub OAuth login loop from a Codespace that breaks API key creation, a refused Hotmail signup, no visible create button. 3 find the models themselves wanting: distillation and quantization for speed, Arabic output degrading against the open-weights original, an extraction returned with confidence 1.0 that was wrong, and one verdict that Groq cannot carry heavy work.

Smaller: 3 hit quota 429s and fall back to a local model, 2 hit client SDK surprises (bad keys accepted, a missing key crashing at startup, strict output failing silently), and 2 cannot predict what long context will cost.

  • Developer plans and billing disabled5
  • Down, out of capacity, or missing models4
  • Can't sign up or get an API key4
  • Model quality drops4
  • Free-tier limits bite3
  • Client SDK quirks bite developers3
See 5 of 25 posts
ᛗᚨᚱᚴᚢᛋ@guitaripod

replaced flaccy's backend model with haiku-5.5. it's so good. i was using llama-3.3:70b-instruct through Groq, but it's been down a few days. no idea why, but i dont even care cause haiku is that good. extremely fast, quality results (opus-4.8 tier), and costs nothing.

UN©️ENSORED 𝕏Men@GutNews247

.@GroqLLC I found a login loop bug! GitHub OAuth from Codespace → /authenticate?token_type=discovery → Logging in... → CSP violation Refused to connect... → back to "Create an account or login" infinite loop. Breaks API key creation. Happy to share full repro + video. Do you have a swag / credits bounty for bug reports? Would love to be considered 🙏

Mazen — sa/acc@ma7dev

to give them the benefits of the doubt they do good stuff but they sometimes ruin the models themselves like the model in english text is acting normally but when trying with arabic text, it just pretend it doesnt comprehend language compared to actual oss model so better eval models to know if they still work for ur use cases cuz fast + bad isnt a good combo especially for systems that should rely on reliability

Max@mosessmax

thanks chief, the cost is mostly on infra not model, most tokens go to email triage, most of it runs on groq’s free tier, i also fall back to gemma4 on my mac because of rate limiting from groq, although i'm also testing jev against my current setup.

5 of the 25 posts behind this answer, one per person. The other 20, and every new one as it’s posted, come with tracking. Get early access

Alternatives· 24 posts from 22 people· written 27 Oct

What are the alternatives to Groq?

Cerebras is what people weigh Groq against — 8 people compare the two fast-inference options, some running both — and every other alternative comes from one or two people.

The Cerebras comparisons come down to the shape of the free offer: Groq's 200,000 tokens a day as a standing tier against Cerebras's 1,000,000 a day drawn from a $5 grant that expires 30 days after it lands — one reader summarises it as one lasting forever and the other lasting a month. Others argue architecture: wafer-scale silicon against many small chips, 44GB of SRAM on a wafer against hundreds of MB per Groq chip.

The rest is thin. 3 weigh Gemini and Google AI Studio, one switching scripts over after Groq scored better on accuracy, another leaving Google AI's stock answers for Groq. 2 found NVIDIA's free NIM endpoints slow by comparison, 17 seconds against Groq's 2. 2 route through OpenRouter and pin Groq first for latency; 2 name Fireworks as another cheap host of open models, 2 benchmark their own self-hosted deployment, 2 weigh Claude, and one person replaced Wispr Flow with their own Groq Whisper dictation at "a fraction of the cost".

  • Cerebras8
  • Gemini3
  • OpenRouter3
  • NVIDIA NIM3
  • Fireworks2
  • Local and self-hosted models2
See 5 of 24 posts
black panther@pantherbllack

1,200,000 FREE TOKENS A DAY ARE SITTING AT TWO PROVIDERS AND MOST PEOPLE KNOW NEITHER [where the 1.2M comes from] → Groq: 200,000 tokens a day, 1,000 requests, 30 a minute. a standing free tier, and they don't train on what you send link → Cerebras: 1,000,000 tokens a day, but out of a $5 grant that dies 30 days after it lands. needs a verified card, doesn't train link the first one lasts forever, the second one lasts a month. together that's 1.2M tokens a day until the clock runs out [4 more while you're there] → Cloudflare Workers AI: 10,000 neurons a day, 300 requests a minute link → OpenRouter: models ending in :free, 50 requests a day, or 1,000 once you've bought 10 credits link → Cohere: 1,000 calls a month link → Alibaba Model Studio: 1M tokens per model, 90 days, Singapore region only link [the trap nobody reads] Alibaba switches you to pay-as-you-go the moment the quota runs out, unless you turn on "Free Quota Only". it's off by default Google's unpaid quota (link trains on what you send. Mistral's free mode (link does too, by default, with an opt-out in the admin panel Groq, Cloudflare and Cerebras don't [how to wire it] → make Groq the default. it doesn't expire and doesn't train on your code → add Cerebras for heavy days while the grant is alive → keep OpenRouter as the fallback when you hit a wall → never send client code to a tier that trains on it you're not saving money on AI. you just never collected what's already free ↓

Artur Meyster@ArturMeyster

Cerebras and Groq are both inference accelerators (I’ve been an early investor in both since 2024), but the architectures are very different. $CBRS is wafer-scale, so inference runs on one giant chip instead of across many small chips wired together. Staying on-chip is much faster than shuttling data between dies. Both use SRAM, but Cerebras has 44GB on a single wafer vs. hundreds of MB per Groq chip. The biggest pushback is that you can’t fit trillion-parameter models on a wafer. But Cerebras has already shown it can split large OpenAI models across multiple wafers and still deliver record speeds. On #2, you’re right that they’re forced to stand up their own infrastructure. They’re innovating on the chip, the racks, and the co-packaging, and there’s no precedent for any of it. In my view that’s an advantage, because nobody in the GPU buildout is doing this. And Cerebras isn’t a pre-commercialization startup like some of the other ASIC players. It has been selling systems and serving inference from its own cloud to large customers for years. Now it’s about scaling the supply chain and getting the data center shells ready, which is well underway. I expect OpenAI revenue to start showing up in Q4 and Q1 2027.

Vicarioy | ML Reasearcher@vicarioy_

Turns out it wasn’t random. OpenRouter rotates requests across upstream providers. I logged every call’s actual serving provider and ran 12 identical back-to-back requests: Groq averaged 0.35s. Novita averaged 1.06s. DeepInfra averaged 3.78s, for the exact same model, same prompt

👾Tushar Bisht👾@tushar_bisht_

for my prorotypes (and hoping to turn one to a product) I search for a faster inference that is free. I tried @nvidia endpoints; the inferences take a hell lot of time (17 sec to parse and transform an email), was recommended @GroqLLC wooped at 2 secs. any recommendations?

5 of the 24 posts behind this answer, one per person. The other 19, and every new one as it’s posted, come with tracking. Get early access

Questions· 6 posts· written 27 Oct

What do people ask about Groq?

Nearly all the questions are about access rather than capability: 4 people ask when paid plans and pay-as-you-go billing come back, some after months of waiting, and 2 ask whether a fast endpoint is really running on Groq's LPU hardware.

The plan question recurs: when developer plans will be activated again, when the dev plan opens, when billing will be enabled again. The tone is more waiting than anger — one says they are "waiting on that for months", another asks to at least be allowed onto pay-as-you-go, and a third says they cannot wait to build with the models once billing returns.

The hardware questions come out of launches. One asks whether an ultrafast tier is served by Nvidia and therefore Groq, reading the change in tokens per second as the explanation; another asks directly whether a new model is running on Groq LPU/X hardware.

  • When will paid plans return?4
  • Is this running on Groq hardware?2
See 5 of 6 posts

5 of the 6 posts behind this answer, one per person. The other 1, and every new one as it’s posted, come with tracking. Get early access

Wishes· 7 posts· written 27 Oct

What do people want from Groq?

Wishes split between more models — 3 people ask Groq to host something they are testing — and keeping the free tier open, which 2 ask for while using it for real work.

The model requests are specific but don't converge: Mimo v2.6, decision models ("already fast and cost effective so they are perfect match for Groq stack"), and a plain "please add support for this model" about something the poster is trying. No single missing model has more than one person asking.

The free-tier wish comes from people already leaning on it: one with no wallet running every model free at about 200,000 tokens a day and shipping agents at zero cost, another who hits token limits and still hopes Groq keeps "the door open" for explorers.

Two one-off asks remain: Groq as an official provider inside Hermes, alongside Cerebras and Fireworks, and letting other agents ping Groq bots directly as a basic requirement of agent workflows.

  • Add support for more models3
  • Keep the free tier open2
  • Official provider in Hermes1
  • Let agents reach Groq bots1
See 5 of 7 posts
innovationbyai@itsrasoull

i have no wallet, so every model i run is free. groq gives me ~200k tokens/day on a 27b model — no credit card, just an api key. you're paying $20/mo to argue with a chatbot. i shipped 3 agents this month on $0. what are you overpaying for?

5 of the 7 posts behind this answer, one per person. The other 2, and every new one as it’s posted, come with tracking. Get early access

In their words

How do people talk about Groq?

The phrases people use most, counted once per person.

  1. 1“free tier”10
  2. 2“api key”5
  3. 3“groq api”4
  4. 4“developer tier”2
  5. 5“token limits”2
  6. 6“fast enough”2
  7. 7“whisper large”2
  8. 8“inference providers”2

Moments· 2 moments

What set off the biggest conversations about Groq?

The days the conversation jumped, and why.

  1. Oct 22, 2026 · 3 posts

    Is Groq NVIDIA's now? Claims and a correction

    Replies state flatly that Groq belongs to NVIDIA, while another post calls that a factual error: Groq remains independent and NVIDIA licensed its technology.

    See the 3 posts
    B@QuantumTumbler

    The case for specialized AI hardware is worth taking seriously. But this a16z essay needs some context. a16z led Gimlet Labs’ $300M funding round last month. The author is Gimlet’s CEO, and the essay argues for the multi-silicon approach his company sells. The claimed 2–10× performance gains aren’t accompanied by enough benchmark detail to reproduce them independently. There’s also a factual error: Groq remains an independent company. NVIDIA licensed its technology; it didn’t rename Groq to “NVIDIA’s LPU.” Heterogeneous inference may offer real efficiency gains. Let’s distinguish the engineering case from the commercial pitch.

  2. Oct 15, 2026 · 2 posts

    Muse on GroqCloud brings LPU questions

    Posts about running Meta's Muse on GroqCloud at 350+ tokens per second were followed by a direct question about whether it runs on Groq LPU/X hardware.

    See the 2 posts

LPU hardware· 9 posts· written 27 Oct

What hardware does Groq run on?

The hardware argument runs in the LPU's favour — 5 people say LPU inference is faster, cheaper or lower-power than GPUs — while 4 point out that small on-chip SRAM caps model size and long-context work.

The bullish case is about economics: low-latency ASICs at 5–10% of the inference market and projected toward 40%, Groq undercutting on prompt-cache pricing at $0.05 per million, tens of GPUs against 2,000-plus LPUs for the same model, and one poster arguing that with Groq LPUs NVIDIA can deliver inference "35x cheaper than Blackwell" today. As models improve, the value of very high tokens per second rises with them.

The limit is memory. Groq trades SRAM for latency, which suits small batches but constrains long context, and discussion of a move to 3D stacked DRAM is framed as breaking that capacity ceiling — though one reply notes that if Groq makes that move, memory suppliers rather than the compute side are the winners. Cerebras's 44GB of SRAM on a single wafer is offered as the contrast.

  • LPUs beat GPUs on speed and power5
  • Small SRAM limits big models4
See 5 of 9 posts
P Equity Research 📰@pequityresearch

Expert Takeaways: Inference > GPU Rental Rates Up ~20%: Driven by HBM supply constraints. Neo-cloud providers are passing hardware cost increases straight to tenants via shorter 1–3 year contracts rather than taking margin cuts. > HBM4 Bottlenecks Capping NVIDIA Vera: NVIDIA’s Vera GPU faces manufacturing constraints—HBM4 issues restrict output to 100k–200k units vs. its 400k–600k theoretical capacity. > Low-Latency ASICs Gaining Share: Agentic AI is splitting workloads into general vs. ultra-fast inference. Low-latency hardware currently sits at 5–10% market share and is projected to reach ~40% long-term. > Etched Hits ~60k Tokens/sec: By stripping general-compute circuitry, Etched boosts transformer utilization to ~90%. It has raised $10B in fresh capital with deployments at Jane Street & JPMorgan. > Groq Commercial Strategy: Offering 50% discounts on cached prompts and batch APIs. LPX service baseline priced at $0.05–$0.09 per million input tokens, with native integration for Rubin. > Groq LPU Benchmarks across Kimi Models: Kimi 1.5 reaches ~1,000 tokens/s, Kimi K2 delivers 400–450 tokens/s, and multi-LPU Kimi 3 hits ~250 tokens/s. > Cloud Benchmark Breakdown (Kimi K2.5): Together AI hits 450 tokens/s (1.37s latency), Fireworks AI leads on speed with sub-second latency (354 tokens/s), and standard neo-clouds trail at 338 tokens/s (1.86s latency). $NBIS $NVDA $MU $CBRS

Nick Dorsey@Midnight_Captl

I can’t get it out of my head that the better these models are getting, the faster tps I want I hadn’t considered before now that the value of tps (tokens per second) may be non linearly correlated with model capability It may be the case that the value of @cerebras / @GroqLLC hardware soars as a result of this. As models are improving rapidly the value of higher tps is going up even more rapidly It’s not inconceivable to me that people would be willing to pay an enormous premium for these ultrafast tokens Post Astra models running at 4k tps is such a huge competitive advantage that it’s worth almost anything to acquire it

Shaw@shawmakesmagic

NVIDIA has literally no competitors who can scale up fast enough to meet demand over the next 5 years, and with Groq LPUs they can deliver post training and inference 35x cheaper than Blackwell *today* on VR with Groq and they are moving to scaled up Feynman now. Even government intervention can’t scale the supply chain that fast

星星之火可以燎原@fireandstart

这篇拆解把“每瓦 token”背后的系统取舍说清楚了：交互式推理受内存带宽和数据搬运影响，未必需要照搬训练芯片的算力配置；HBM4 带来的带宽提升、减少为训练优化而闲置的资源，才是效率故事的重要部分。Maia 300 与 Jalapeño 的类比也提醒我们，芯片规格只是起点，最终还要看软件栈能否把硬件潜力兑现。相较之下，Groq、Cerebras 用 SRAM 换低延迟，适合小批量场景，却会受容量和长上下文限制。所谓每瓦多出约 40% token，值得关注，但仍需在相同模型、精度、批量、上下文长度和端到端延迟条件下比较。推理专用硬件能否成为主流，关键不只是一张 FLOPs/W 图，而是整个服务负载下稳定、可复现的成本与吞吐。

Techintosh@userHandle

Groq評估以3D堆疊DRAM突破SRAM容量限制 外媒報導，近期遭輝達實質性收購的Groq，正評估下一代晶片採用3D堆疊DRAM封裝（3DS）。重點不只是把記憶體堆高，而是讓超大型AI模型推論有更多容量可用。 Groq現有LPU架構以高速推論聞名，但SRAM容量有限；當單一晶片要面對千億級參數模型時，可能遭遇記憶體不足的限制。DRAM原生讀寫速度略低於SRAM，卻能提供更大的容量，成為複雜模型部署時的重要取捨。 3DS的作法是把多個半導體裸晶垂直堆疊，並以TSV（矽通孔）在DRAM晶片內建立數千個垂直互連，藉此提高單位面積的記憶體密度。相較同樣使用TSV堆疊的HBM主要追求極高資料交換速率，3DS更偏向擴充容量，同時維持既有DRAM標準介面協定。 若這項評估最終落地，代表AI推論晶片的競爭焦點可能不只在運算速度與頻寬，也更取決於能否在有限面積內容納足夠模型參數。對需要載入大型模型的推論工作負載而言，容量或許會成為和速度同等關鍵的設計條件。 你認為大型AI推論晶片下一步最該優先解決的，是記憶體容量、頻寬，還是成本？

5 of the 9 posts behind this answer, one per person. The other 4, and every new one as it’s posted, come with tracking. Get early access

NVIDIA deal· 7 posts· written 27 Oct

Did NVIDIA buy Groq?

People split on it: 4 say or assume NVIDIA bought Groq, while 3 correct that Groq stays independent and NVIDIA only licensed the technology.

The takeover version is stated flatly rather than argued — that NVIDIA bought Groq after cheap inference, that "groq es de nvidia", and that NVIDIA "had to buy Groq" after failing to buy Cerebras quietly. Those posts treat it as settled background in wider threads about GPUs and investors.

The correction is equally direct. One reply says NVIDIA tried to buy Cerebras, was declined, and signed a deal with Groq instead. A longer post calls the acquisition claim a factual error: Groq remains an independent company and NVIDIA licensed its technology rather than rebranding the LPU. A Chinese-language post describes a $20 billion technology licensing agreement and NVIDIA's parallel investment in d-Matrix to run competitor chips alongside its GPUs.

Nothing in the posts shows anyone changing their mind; the two readings sit side by side.

  • NVIDIA bought Groq4
  • A licensing deal, not a takeover3
See 5 of 7 posts
Max Kalashnikov@maxkalashnikoff

Yeah. For me Nvidia’s domination in GPUs is the most puzzling thing. Ok, Google did TPU.. at best it powers Gemini. Groq had cheap inference - Nvidia bought them. AMD - not sure what are they doing. In theory, chips only for inference should be easier and cheaper. But no, all need Nvidia’s GPUs

B@QuantumTumbler

The case for specialized AI hardware is worth taking seriously. But this a16z essay needs some context. a16z led Gimlet Labs’ $300M funding round last month. The author is Gimlet’s CEO, and the essay argues for the multi-silicon approach his company sells. The claimed 2–10× performance gains aren’t accompanied by enough benchmark detail to reproduce them independently. There’s also a factual error: Groq remains an independent company. NVIDIA licensed its technology; it didn’t rename Groq to “NVIDIA’s LPU.” Heterogeneous inference may offer real efficiency gains. Let’s distinguish the engineering case from the commercial pitch.

5 of the 7 posts behind this answer, one per person. The other 2, and every new one as it’s posted, come with tracking. Get early access

Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.

Also in Inference and model APIs

Get every post about Groq, filtered and labelled

This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.