Groq
AI infrastructure company providing fast, predictable generative-AI inference with its Language Processing Unit (LPU).
Website: https://groq.com/
Cover Block
Publicly reported
| Field | Detail |
|---|---|
| Name | Groq |
| Tagline | AI infrastructure company providing fast, predictable generative-AI inference with its Language Processing Unit (LPU) [Groq, retrieved 2026] |
| Headquarters | Mountain View, California [TechCrunch, September 2018] |
| Founded | 2016 [TechCrunch, September 2018] |
| Stage | Series D+ [TechCrunch, June 2026] |
| Business model | API / Developer Platform [TechCrunch, August 2024] |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | Global / Remote-First [TechCrunch, August 2026] |
| Growth profile | Venture Scale |
| Founding team | Co-Founders (2): Jonathan Ross, Douglas Wightman [TechCrunch, September 2018] |
| Funding label | $100M+ |
| Total disclosed funding | Approximately $2,742,300,000 (sum of disclosed rounds) [TechCrunch, September 2018] [TechCrunch, August 2024] [TechCrunch, September 2025] [TechCrunch, June 2026] [TechCrunch, August 2026] |
Links
Publicly reported
- Website: https://groq.com/
- X / Twitter: https://wow.groq.com/careers
Summary and Signal
PUBLIC Groq builds AI inference infrastructure, selling access to custom chips and a cloud API designed for fast, predictable generative-AI serving, and it warrants attention now because the company has paired unusually large financing with a visible shift from chip maker to inference-cloud operator [TechCrunch, August 2024] [TechCrunch, August 2026]. Founded in 2016 in Mountain View by Jonathan Ross and Douglas Wightman, Groq emerged from Google-adjacent technical roots: Ross previously worked on Google’s TPU effort, while Wightman had engineering experience at Google and Alphabet’s X [TechCrunch, September 2018].
The product case rests on Groq’s Language Processing Unit, or LPU, and the company’s GroqCloud platform, which hosts open models including Meta’s Llama, Google’s Gemma, OpenAI’s Whisper, and Mistral’s Mixtral through an API model [TechCrunch, August 2024] [GroqDocs, retrieved 2026]. Its differentiation claim is lower-latency, more predictable inference relative to general-purpose GPU stacks, although much of that positioning remains company-described and should be treated cautiously until backed by more third-party benchmarks [Groq, retrieved 2026].
On team quality, the public record supports strong technical founding credentials, and later leadership additions have tilted toward operating scale: Stuart Pann joined as COO in 2024, and Yann LeCun was announced as a technical adviser alongside the 2024 financing [TechCrunch, August 2024]. Leadership continuity is a point to watch, however, because reporting after Nvidia’s December 2025 licensing and talent transaction indicates a post-Ross transition to Adam Winter as chief executive [Reuters, retrieved 2026] [The Daily Star, retrieved 2026].
Capital formation has been a defining part of the story. Groq has disclosed roughly $2.74 billion in financing across rounds that include $640 million in August 2024, $750 million in September 2025, $650 million in June 2026, and $350 million in August 2026, with BlackRock and Disruptive among the recurring lead investors [TechCrunch, August 2024] [TechCrunch, September 2025] [TechCrunch, June 2026] [Bloomberg, August 2026]. The business model appears increasingly bifurcated between proprietary infrastructure and usage-based cloud services, which can be attractive if utilization stays high, but also exposes the company to execution risk in data-center buildout, pricing discipline, and customer concentration that public sources do not yet clarify [GroqDocs, retrieved 2026] [TechCrunch, August 2026].
Over the next 12 to 18 months, the key question is whether Groq can convert headline financing, a 13-data-center footprint, and developer adoption claims into durable cloud revenue without losing strategic coherence after the Nvidia deal and valuation reset reported in 2026 [TechCrunch, August 2026] [Reuters, retrieved 2026] [Bloomberg, August 2026]. The company is clearly relevant in AI infrastructure; the remaining work for investors is to separate architectural promise from operating proof.
One source, partially checked -- Relies primarily on TechCrunch, Bloomberg, Reuters, and company materials; several product and adoption claims remain company-described or only partially corroborated.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Series D+ |
| Business Model | API / Developer Platform |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | Global / Remote-First |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2) |
| Funding | $100M+ total disclosed, approximately $2,742,300,000 |
Company Overview
PUBLIC
Groq traces back to 2016 in Mountain View, California, with a founding team led by Jonathan Ross and Douglas Wightman [Crunchbase, retrieved 2026] [Groq, retrieved 2026]. The plain description is still the useful one: this is an AI infrastructure company building custom inference chips and cloud access around its Language Processing Unit, or LPU, with the company now presenting itself as a neocloud for fast inference [Groq, retrieved 2026].
The chronology that can be established from company and database records is relatively clean. Crunchbase records Groq as founded in 2016 and headquartered in Mountain View, while the company website describes a business that has moved from chip development toward a broader inference cloud platform over time [Crunchbase, retrieved 2026] [Groq, retrieved 2026]. One visible milestone in that expansion came in March 2024, when Groq announced Groq Systems and the acquisition of Definitive Intelligence, adding a business applications layer to the infrastructure stack [Groq, retrieved 2026].
Public capital formation also marks the company's path. Crunchbase lists Groq in a late-stage venture posture, and the company website highlights a $650 million financing in 2026 tied to scaling its AI inference cloud business [Crunchbase, retrieved 2026] [Groq, retrieved 2026]. Taken together, the public record supports a company that began as a semiconductor bet and has since widened into a full-stack inference platform, although the corporate and operating structure beyond that high-level shift is only partially documented in the cited public materials [Groq, retrieved 2026] [Crunchbase, retrieved 2026].
Well sourced -- Confirmed by Crunchbase and the company website.
The Product and the Stack
Public record plus analysis
Groq’s product story is no longer only about a custom chip. Publicly, the company is presenting a combined stack: purpose-built AI inference hardware, cloud delivery, and model access through GroqCloud, a developer platform that exposes hosted models over API endpoints [TechCrunch, August 2024] [GroqDocs, retrieved 2026]. The technical claim that matters most is not general model quality but serving behavior, specifically fast and relatively predictable inference from its Language Processing Unit, or LPU, rather than from general-purpose GPUs [Groq, retrieved 2026] [TechCrunch, September 2018]. That distinction is central to how Groq is marketed and to where it fits in the stack, though the strongest performance language still comes from the company itself and should be read as company positioning rather than independently benchmarked proof [Groq, retrieved 2026].
The commercial surface is clearer than the underlying architecture. TechCrunch reported in August 2024 that GroqCloud hosted Meta’s Llama 3.1, Google’s Gemma, OpenAI’s Whisper, and Mistral’s Mixtral, which places Groq less as a frontier-model lab and more as an inference layer for open and widely used models [TechCrunch, August 2024]. The company’s documentation also shows a Models API endpoint and commercial terms built around cloud services and AI model services, with pricing either posted on the website or specified in an order form [GroqDocs, retrieved 2026]. By 2026, Groq was also describing LPX as a product that works alongside Nvidia’s next-generation GPUs, which suggests a more pragmatic interoperability stance than a pure replacement pitch, although that specific framing is company-sourced [Groq, retrieved 2026].
The product footprint has expanded beyond raw inference serving, but the evidence is mixed on how far that expansion goes in practice. In March 2024, Groq formed Groq Systems and acquired Definitive Intelligence, a Palo Alto company building business AI applications such as chatbots, analytics tools, and documentation tools, indicating some interest in moving up the stack toward enterprise workflows [TechCrunch, March 2024]. A separate 2022 acquisition of Maxeler Technologies is commonly cited as adding dataflow-systems expertise, though the available support here is weaker and rests on Wikipedia rather than primary company disclosure [Wikipedia, retrieved 2026]. On balance, the public record supports a view of Groq as an inference infrastructure company with a growing software and services layer, but not yet enough verified detail to map the full control plane, orchestration layer, or enterprise feature set with confidence.
No independent source found -- This section combines independent reporting from TechCrunch with company website and documentation claims that are material to the product description but not fully corroborated by independent technical benchmarks.
The Market They Are Entering
PUBLIC The market matters now because inference has shifted from a technical afterthought to the cost, latency, and availability layer that determines whether generative AI reaches production at scale, and Groq is explicitly positioning itself around that bottleneck rather than around model creation alone [Groq, retrieved 2026] [TechCrunch, August 2026].
Public source support for formal TAM, SAM, or SOM is thin in the materials reviewed, so the safer frame is to treat Groq as operating at the intersection of AI semiconductors, cloud infrastructure, and API-based model serving, with sizing evidence limited to analogous public market signals rather than a directly cited market study [TechCrunch, August 2024] [TechCrunch, August 2026]. What is verifiable is that the company has moved from being described primarily as an AI chip startup in 2024 to a neocloud provider by 2026, which implies its addressable market expanded from hardware procurement into recurring inference consumption by developers, enterprises, and AI-native companies [TechCrunch, August 2024] [TechCrunch, August 2026]. That distinction matters for investors because infrastructure spend tied to API usage can compound with model adoption in a different way than one-time accelerator sales, but the evidence here supports the category shift more clearly than it supports a quantified market share claim [GroqDocs, retrieved 2026] [TechCrunch, August 2026].
Demand drivers in the cited research are straightforward. Groq's public wedge is fast, predictable inference on purpose-built LPUs, and that pitch maps to a broad market concern that only intensifies as generative AI moves into customer-facing and operational workloads where response time is visible and compute cost scales with usage [Groq, retrieved 2026] [TechCrunch, September 2018]. The company's own materials state that millions of developers run trillions of tokens on Groq every week, and GroqCloud separately claims more than 2 million developers are building on the platform, but those figures remain company-reported and should be read as directional rather than independently verified demand proof [Groq, retrieved 2026] [GroqCloud, retrieved 2026]. More grounded external evidence comes from distribution and infrastructure expansion: TechCrunch reported model hosting for Llama, Gemma, Whisper, and Mixtral in 2024, and reported 13 data centers across North America, Europe, the Middle East, and Asia-Pacific by August 2026, both of which suggest Groq is serving a market that increasingly values geographic coverage and model breadth alongside raw chip performance [TechCrunch, August 2024] [TechCrunch, August 2026].
The closest adjacent markets are hyperscale GPU cloud, AI inference APIs, and enterprise AI application infrastructure. Groq's offering can substitute for portions of Nvidia-centric inference stacks when buyers prioritize latency determinism or token economics, but it can also sit alongside incumbent hardware, especially because the company says LPX works with Nvidia's next-generation GPUs and Reuters reported a non-exclusive licensing arrangement between the two companies in late 2025 [Groq, retrieved 2026] [Reuters, retrieved 2026]. That creates a more nuanced market position than a simple chip-versus-chip contest. In practice, Groq appears exposed not only to semiconductor capital cycles, but also to cloud platform adoption, open-model distribution, and application-layer demand for hosted inference [TechCrunch, August 2024] [TechCrunch, June 2026].
Macro and regulatory forces cut both ways. On the positive side, sovereign and enterprise interest in AI capacity appears to favor providers that can stand up data-center footprint across regions, and Groq's reported 13-site network by August 2026 suggests management is building for that requirement rather than for a single-region developer market [TechCrunch, August 2026]. On the cautionary side, this remains a capital-intensive market where infrastructure buildout, chip supply, and strategic partnerships can reshape competitive positioning quickly; the valuation swing from $6.9 billion post-money in September 2025 to $3.5 billion in August 2026 underlines how sensitive the category is to financing conditions and strategic transactions, even when demand for inference remains intact [TechCrunch, September 2025] [Bloomberg, August 2026].
| Segment or signal | What public sources support | Why it matters |
|---|---|---|
| AI chip startup to neocloud transition | Groq was described as an AI chip startup in 2024 and a neocloud provider in 2026 [TechCrunch, August 2024] [TechCrunch, August 2026] | Expands the relevant market from hardware sales into recurring inference services |
| Geographic infrastructure footprint | 13 data centers across North America, Europe, the Middle East, and Asia-Pacific by August 2026 [TechCrunch, August 2026] | Suggests demand is becoming global and capacity-constrained rather than purely experimental |
| Developer and token usage claims | Millions of developers and trillions of tokens weekly, according to Groq; 2M+ developers on GroqCloud, according to company materials [Groq, retrieved 2026] [GroqCloud, retrieved 2026] | Indicates broad top-of-funnel interest, though still company-reported |
| Strategic overlap with Nvidia ecosystem | Non-exclusive licensing arrangement with Nvidia in December 2025 [Reuters, retrieved 2026] | Points to coexistence with incumbent infrastructure, not only direct displacement |
The evidence supports a market that is broadening faster than it is stabilizing. Groq's opportunity appears tied less to a neatly bounded chip TAM than to the growing spend pool around production inference, but the public record here is stronger on category direction than on third-party market sizing.
One source, partially checked -- This section relies mainly on TechCrunch and Reuters, with supplemental company-reported product and usage claims that are not independently verified.
The Competitive Field
Market map
MIXED Groq sits in a narrow but consequential lane between incumbent GPU platforms and newer inference-focused clouds, with its case resting on whether purpose-built inference hardware can win a durable share of production AI workloads from more general compute stacks [TechCrunch, August 2024] [Groq, retrieved 2026].
| Company | Positioning | Stage / Funding | Notable Differentiator | Source |
|---|---|---|---|---|
| Groq | AI inference infrastructure and developer API built around its Language Processing Unit and GroqCloud | Series D+; disclosed funding of about $2.74B | Purpose-built inference architecture positioned around fast, predictable serving rather than general-purpose GPU compute | [TechCrunch, August 2024] [TechCrunch, September 2025] [TechCrunch, June 2026] [TechCrunch, August 2026] |
| Nvidia | Incumbent AI compute platform spanning chips, systems, and software | Public company | Scale, distribution, and software ecosystem breadth; also entered a non-exclusive licensing arrangement for Groq technology in 2025 | [Reuters] [TechCrunch, June 2026] |
The competitive map is easiest to read in three layers. At the top sits Nvidia, which remains the incumbent standard for AI compute because it combines chips, systems, and an entrenched software stack, not just silicon [Reuters] [TechCrunch, August 2024]. Groq is the challenger in inference-specific infrastructure, arguing that many production workloads care less about training flexibility and more about latency, predictability, and operating efficiency at scale [TechCrunch, August 2024] [Groq, retrieved 2026]. Adjacent substitutes sit one step to the side rather than head-on: enterprises can still choose to serve open models through general cloud infrastructure or standard GPU fleets, which makes the real alternative to Groq not only a rival chip vendor but also the status quo of staying on broad-purpose compute [TechCrunch, August 2024].
Groq's defensible edge today appears to be a combination of architecture focus and capital access, but both edges need qualification. The product story is unusually coherent for an infrastructure company: custom inference chips, a cloud API, and a growing neocloud positioning are all pointed at the same bottleneck, which is production inference rather than model training [TechCrunch, August 2024] [TechCrunch, August 2026]. The company has also raised large amounts of capital across 2024 to 2026 and said by August 2026 that it operated 13 data centers across North America, Europe, the Middle East, and Asia-Pacific, which matters because developers will not trust a specialized platform if capacity is thin or geographically narrow [TechCrunch, August 2026]. That said, the durability of this edge is still perishable. Speed and predictability are valuable only if they remain meaningfully better than what customers can get from incumbent GPU-based offerings, and Groq's biggest proof points on developer scale come primarily from company channels rather than independent disclosures [Groq, retrieved 2026] [GroqCloud, retrieved 2026].
The main exposure is straightforward: Nvidia owns the broader compute standard and can compete from a position of channel control. Even where Groq is differentiated, Nvidia can answer with distribution, developer familiarity, and a much larger installed base across enterprises and cloud providers [Reuters] [TechCrunch, August 2024]. Groq is also exposed to category boundaries of its own making. Its identity is increasingly tied to inference cloud services, which sharpens the message but also narrows the field of battle; if customers want one vendor across training, fine-tuning, orchestration, and inference, the integrated generalist platform can still be the simpler buying decision [TechCrunch, August 2026] [GroqDocs, retrieved 2026]. The December 2025 licensing arrangement with Nvidia cuts both ways on that point. It validated the underlying technology, according to Reuters' reporting on the transaction, but it also underscored that the incumbent has the balance sheet and ecosystem reach to absorb useful innovation without ceding category control [Reuters] [TechCrunch, June 2026].
Over the next 18 months, the most plausible competitive scenario is a bifurcation rather than a winner-take-all outcome. Nvidia is the likely winner if enterprise buyers continue to consolidate vendors around a familiar full-stack platform, especially if inference performance gains become easier to replicate inside GPU-centered offerings [Reuters] [TechCrunch, August 2024]. Groq is the likely winner if inference becomes a distinct procurement category, where buyers select for latency, determinism, and usage economics independently from training infrastructure, and if its neocloud footprint converts developer usage into durable enterprise workloads [TechCrunch, August 2026] [GroqDocs, retrieved 2026]. The loser if that latter transition stalls is not necessarily Groq's technology, which already drew strategic interest, but Groq's ambition to stand alone as a scaled cloud platform rather than an enabling architecture inside someone else's stack [Reuters] [TechCrunch, June 2026].
One source, partially checked -- This section relies on TechCrunch and Reuters for independent reporting on Groq's funding, product positioning, Nvidia relationship, and infrastructure footprint, with company materials used for product claims that are not fully independently corroborated.
Opportunity
PUBLIC
The prize here is unusually large: if Groq can turn a differentiated inference stack into the default service layer for latency-sensitive generative AI, it could graduate from a chip startup into a globally relevant AI utility with strategic value well beyond component economics [TechCrunch, August 2024] [TechCrunch, August 2026].
The headline opportunity is not simply selling another accelerator. The more interesting path is owning the inference layer that developers and enterprises touch every day, with proprietary silicon underneath and a cloud API on top. Public evidence suggests that Groq has already moved in that direction: TechCrunch described the company in 2026 as pivoting from primarily an AI-chip company toward a neocloud provider, while its GroqCloud platform hosts widely used open models such as Llama, Gemma, Whisper, and Mixtral [TechCrunch, August 2026] [TechCrunch, August 2024]. That matters because infrastructure companies tend to become more valuable when customers buy ongoing service consumption rather than one-off hardware, and Groq's footprint, 13 data centers across North America, Europe, the Middle East, and Asia-Pacific by August 2026 according to TechCrunch, gives at least some public evidence that the delivery layer is becoming real rather than conceptual [TechCrunch, August 2026].
The pathways to scale are visible, even if the evidence remains uneven on end-market adoption and economics.
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Neocloud default for open-model inference | Groq becomes a preferred API for developers and AI-native companies that need low-latency access to open models, with usage expanding as more applications move from experimentation to production | The product already exposes hosted models through GroqCloud, and the company has continued financing the cloud buildout through large rounds in 2024 to 2026 [TechCrunch, August 2024] [TechCrunch, June 2026] [TechCrunch, August 2026] | The company offers an integrated developer platform rather than only chips, and management has publicly positioned speed and predictability as the wedge for inference workloads [Groq, retrieved 2026] [TechCrunch, August 2024] |
| Strategic infrastructure layer beside Nvidia | Groq wins as a complement to, not a replacement for, incumbent GPU stacks, supplying inference where determinism and throughput matter most | Nvidia signed a non-exclusive licensing arrangement for Groq technology in December 2025, and Groq has said LPX works alongside Nvidia GPUs [Reuters] [Groq, retrieved 2026] | A coexistence model is more plausible than full displacement because enterprises already run Nvidia-heavy environments and may adopt specialized inference hardware incrementally [Reuters] [TechCrunch, June 2026] |
| Public-sector and enterprise channel expansion | Groq converts distribution partnerships and application-layer packaging into larger enterprise deployments, especially where procurement favors approved channels and integrated solutions | TechCrunch reported a Carahsoft partnership for public-sector sales and Groq's formation of Groq Systems through the Definitive Intelligence acquisition in March 2024 [TechCrunch, August 2024] [TechCrunch, March 2024] | The company is no longer presenting as chip-only. It is adding channel access and business-application capabilities that can shorten the path from infrastructure to budget owner [TechCrunch, March 2024] [TechCrunch, August 2024] |
The compounding mechanism would come from stacking three layers that usually sit in separate businesses: silicon, cloud capacity, and developer distribution. If a developer adopts GroqCloud for one hosted model because latency is attractive, that usage can pull through more token volume, justify more data-center buildout, improve platform economics, and make Groq a better launch venue for additional models and enterprise packages. There is some early public evidence that this loop has started. Groq says millions of developers run trillions of tokens on its platform every week, though that claim is company-reported and not independently verified, and TechCrunch separately reported that the company had expanded to 13 data centers by August 2026 [Groq, retrieved 2026] [TechCrunch, August 2026]. Even with conservative treatment of company claims, the strategic logic is straightforward: recurring inference demand is more likely to compound than hardware design wins alone.
The size of the win depends on which business Groq ultimately becomes. A useful public marker is Groq's own financing history: TechCrunch reported a $6.9 billion post-money valuation in September 2025, while Bloomberg reported a $3.5 billion valuation in August 2026 after the Nvidia transaction [TechCrunch, September 2025] [Bloomberg, August 2026]. If the first scenario plays out and Groq emerges as a scaled inference platform rather than a niche chip vendor, an outcome above its prior $6.9 billion mark is easy to imagine on strategic grounds alone, particularly given Nvidia's willingness to license Groq technology and hire key personnel in a transaction Reuters described as worth $17 billion [Reuters] [TechCrunch, June 2026]. A plausible upside frame is therefore low tens of billions of dollars in enterprise value (scenario, not a forecast), not because current public metrics prove that scale, but because the market has already shown willingness to assign multi-billion-dollar value to both Groq's technology and its role in the AI infrastructure stack [TechCrunch, September 2025] [Reuters].
One source, partially checked -- Relies on TechCrunch and Bloomberg for financing, valuation, product evolution, and infrastructure footprint, with some material upside arguments informed by company-reported adoption claims and Groq's own product positioning [TechCrunch, August 2024] [TechCrunch, August 2026] [Bloomberg, August 2026] [Groq, retrieved 2026].
Sources
Publicly reported
[Groq, retrieved 2026] Groq is the premier neocloud for fast inference | https://groq.com/
[TechCrunch, September 2018] Secretive semiconductor startup Groq raises $52M from Social Capital | https://techcrunch.com/2018/09/05/secretive-semiconductor-startup-groq-raises-52m-from-social-capital/
[TechCrunch, June 2026] AI chipmaker Groq confirms $650M raise, re-staffs after Nvidia's $20B not-acqui-hire deal | https://techcrunch.com/2026/06/22/ai-chipmaker-groq-confirms-650m-raise-re-staffs-after-nvidias-20b-not-acqui-hire-deal/
[TechCrunch, August 2024] AI chip startup Groq lands $640M to challenge Nvidia | https://techcrunch.com/2024/08/05/ai-chip-startup-groq-lands-640m-to-challenge-nvidia/
[TechCrunch, September 2025] Nvidia AI chip challenger Groq raises even more than expected, hits $6.9B valuation | https://techcrunch.com/2025/09/17/nvidia-ai-chip-challenger-groq-raises-even-more-than-expected-hits-6-9b-valuation/
[TechCrunch, August 2026] Groq raises $350M to fuel its pivot from AI chips to neocloud | https://techcrunch.com/2026/08/17/groq-raises-350m-to-fuel-its-pivot-from-ai-chips-to-neocloud/
[GroqDocs, retrieved 2026] Groq Services Agreement | https://console.groq.com/docs/legal/services-agreement
[Reuters, retrieved 2026] Groq Inc | https://www.reuters.com/company/groq-inc/
[The Daily Star, retrieved 2026] AI chipmaker Groq raises $650 mln after Nvidia licensed its chips | https://www.thedailystar.net/news/tech-startup/news/ai-chipmaker-groq-raises-650-mln-after-nvidia-licensed-its-chips-4206196
[Bloomberg, August 2026] Groq valued at $3.5 billion in funding round after Nvidia deal | https://www.bloomberg.com/news/articles/2026-08-17/groq-valued-at-3-5-billion-in-funding-round-after-nvidia-deal
[Crunchbase, retrieved 2026] Groq | https://www.crunchbase.com/organization/groq
[Wikipedia, retrieved 2026] Groq - Wikipedia | https://en.wikipedia.org/wiki/Groq
Articles about Groq
- Groq's $3.5 Billion Valuation Anchors a Bet on Predictable AI Inference — The company, which invented the Language Processing Unit, has raised over $2.7 billion to build what it calls a 'neocloud' for low-latency AI model serving.