Nothing on the record argues against opening diligence.
The written sections below reach their own conclusion from the full evidence. Where they and this weighted read differ, the sections carry the argument and the citations.
The Short Answer
What the whole report comes to, and what to do about it.
The revenue trajectory is real and corroborated by independent parties. Fireworks AI's own Series D announcement dated July 15, 2026 claims over $1 billion in annualized revenue run rate; an independent research firm estimated $800 million in annualized revenue as of May 2026, up from approximately $305 million at end of 2025; and the Wall Street Journal reported annual recurring revenue exceeding $280 million at the October 2025 Series C. The growth is steep and corroborated, but no audited financial statement exists in any record searched.
The valuation you would be paying against is $17.5 billion, established by the July 2026 Series D led by Atreides Management, Index Ventures, and TCV — a 4.4× step-up from the $4 billion Series C just nine months earlier. That valuation embeds growth expectations that rest on consumption revenue with no confirmed contractual floor: the billing model is pre-paid credits drawn down per token, not multi-year committed contracts, meaning every named customer — Samsung, Uber, DoorDash, Notion, Shopify, and Upwork — can route traffic elsewhere without breaching an agreement.
The strongest argument against the break-on-close thesis is that Fireworks AI's technical differentiation — its custom inference kernel, speculative decoding stack, and fine-tuning infrastructure — is genuinely hard to replicate, and Nvidia absorbing that engineering capability would neutralize a competitor while acquiring the team that built it. The company's founding team came from PyTorch leadership at Meta and Google, and the platform already appeared at Nvidia GTC in March 2026, suggesting a working relationship exists.
The single most important thing that could not be established is the change-of-control language in the Microsoft Azure Foundry partnership agreement, the Oracle OCI agreement, and the AWS relationship. Those three documents — none of which are on the record — determine whether Nvidia's largest cloud competitors can terminate or renegotiate the distribution arrangements that carry Fireworks AI's inference to a material portion of its customer base the moment an acquisition closes. They would be found in Fireworks AI's contract files and must be the first items requested in any data room.
Before tabling an offer: demand the Microsoft, Oracle, and AWS partnership agreements in full; request the cap table as of the Series D close with liquidation preference multiples for each series; obtain audited or reviewed financials for the trailing twelve months; and ask for the revenue split between committed annual contracts and pure consumption accounts.
Which Company This Is
The entity this report is about, and how we know it is that one and not another of the same name.
Fireworks AI, Inc. trades under the name Fireworks AI and operates through the domain fireworks.ai, which anchors this identification. The company is headquartered in the San Francisco Bay Area — sources variously list Redwood City and San Mateo, California — and was founded in 2022, placing it at roughly four years old as of this report date. It is a private company at Series D stage, having closed a $1.505 billion round in July 2026 at a $17.5 billion post-money valuation, as announced by the company and confirmed by multiple independent sources.
No registry identity was confirmed for this entity. The UK company register returned no match. The global LEI index returned no match. No filing registry entry was located that would supply a company registration number or jurisdiction of incorporation. The legal name "Fireworks AI, Inc." appears in copyright footers across the company's own website and in legal counsel announcements from Orrick, suggesting a US incorporation, almost certainly in Delaware given the "Inc." suffix and the standard practice of venture-backed US startups, but the specific jurisdiction and registration number could not be established from the evidence searched.
One similarly-named entity exists and must be distinguished: Firework (LinkedIn profile at linkedin.com/company/fireworkhq) is a separate company described as a video commerce platform. Nothing from that entity has been attributed to this report's subject.
Identification confidence is high on the trading name, domain, and commercial facts; it is incomplete on formal registry identity, and a Delaware or California corporate registry search would settle the incorporation details definitively.
What They Do, And How The Money Works
The business as it actually operates: what is sold, to whom, on what terms, and where the revenue comes from.
Fireworks AI sells inference-as-a-service and model customization tooling, delivered through a cloud API, to developers and engineering teams building AI-powered products. The core transaction is compute time: a customer sends a prompt to Fireworks AI's infrastructure, the platform runs it through a hosted open-weights model and returns a response, billing the customer per token consumed. That is the atomic unit of the business.
The product page distinguishes three delivery modes. Serverless inference is pay-per-token with no reservation required, suited to variable or early-stage workloads. On-demand dedicated GPU deployment lets customers reserve capacity for predictable, high-volume production traffic. Fine-tuning — both supervised fine-tuning and reinforcement fine-tuning — lets customers adapt a base model to their own data and then serve it on Fireworks infrastructure, with the company's own blog stating fine-tuning is available at no additional cost beyond the inference charges.
The serverless pricing documentation confirms per-token billing with tiered serving paths (Standard, Priority, and Fast). Spot figures from a third-party aggregator show listed rates such as $0.150 per million tokens for a 120-billion-parameter class model and $0.070 for a 20-billion-parameter class model, though these are third-party observations and pricing changes without notice. The billing FAQ describes a pre-paid credit system: customers purchase credits in advance, and usage is deducted against that balance per token for serverless inference and per GPU-hour for dedicated deployments. New accounts receive $1 in free starter credits, approximately one million tokens on a 70-billion-parameter class model. Enterprise contracts exist as a separate tier, with the enterprise page directing buyers to contact sales rather than publishing rates.
The buyer profile visible in the evidence spans a wide range. Named customers at the Series C announcement include Samsung, Uber, DoorDash, Notion, Shopify, and Upwork — large enterprises with substantial engineering organisations. The platform simultaneously serves a self-serve developer base: the Series C announcement noted over 10,000 companies and hundreds of thousands of developers, a tenfold increase from the Series B. An outside customer tracker detected 368 companies using Fireworks AI as of August 2026, though that figure reflects only what can be observed from outside and will undercount actual usage.
The revenue structure is transactional and consumption-based rather than seat-licensed, which means it scales directly with customer AI workload volume and falls if customers reduce usage or switch providers. The Series D announcement states the company crossed $1 billion in annualised revenue run rate as of July 2026, and an independent analyst estimated annualised revenue at $800 million in May 2026, up from approximately $305 million at end of 2025 — a trajectory consistent with consumption growth rather than a fixed subscription base. The Wall Street Journal reported annual recurring revenue exceeding $280 million at the time of the October 2025 Series C, though the applicability of "recurring" to a consumption model is an important caveat: the revenue recurs only as long as customers keep running workloads.
What is not visible from outside is the gross margin on that revenue — GPU compute is the dominant cost, and the split between what Fireworks AI pays for infrastructure and what it retains is not disclosed in any evidence available here.
For an acquirer evaluating this business: Fireworks AI converts GPU compute into per-token API revenue at scale, and the durability of that revenue depends entirely on whether customers are locked in by fine-tuned models and integrations or remain free to route the same tokens to a cheaper provider tomorrow.
Who Runs It, And How To Reach Them
The named people who decide things here, what is known about them, and where a route to them exists.
Lin Qiao is the co-founder and CEO of Fireworks AI and is the person any acquirer would be across the table from. Her LinkedIn profile confirms the role, and her prior position is stated on the company's own team page: she was previously Head of PyTorch at Meta. Business Insider reported in April 2026 that she "helped build PyTorch" at Meta before founding Fireworks AI. The company was founded in 2022 per a private-market company profile, placing her tenure as CEO at roughly four years. She announced the Series D personally on LinkedIn in July 2026, which is consistent with her being the controlling voice on capital decisions.
Benny Chen is listed as a co-founder on the company's own team page, but no title beyond co-founder, no prior employer, and no contact route appear in any of the evidence searched. This is a single-source finding from a page that may not be current.
Chenyu Zhao is identified as a co-founder and described as founding CTO by a third-party executive directory, which states he led machine learning architecture and integrated Google Vertex AI advancements. That source is a single secondary source and the title "founding CTO" may indicate a past rather than current role; no corroboration from a primary source was found.
Steven Wu joined as SVP Finance and Business Operations in February 2026, per his LinkedIn profile. Before Fireworks AI he served as CFO and Chief Business Officer at Guideline from September 2022. He is the most senior finance executive on record and would be a key counterpart in any acquisition negotiation on financial representations.
No Chief Revenue Officer, General Counsel, or Head of Engineering is named with enough confidence in the evidence to include here. The Sequoia job board lists open roles for a Compensation Partner and HR Business Partner, GTM, suggesting the people-operations function is still being built out at scale.
One structural point for this decision: the leadership footprint that is publicly verifiable is narrow. Lin Qiao is clearly the founder-CEO and the face of every major capital event. That concentration is a material diligence item the evidence does not resolve.
The Money: Funding, Valuation, Runway
Every financial figure on the record, dated and sourced, and what the shape of it implies.
Fireworks AI, Inc. has raised capital across four named rounds between 2024 and 2026. The earliest round in the evidence is a Series A of $25 million, closed in March 2024, with no lead investor named in the record. A Series B of $52 million followed in July 2024, led by Sequoia Capital at a post-money valuation of $522 million — one source from the company's own blog puts the post-money figure at $552 million, a minor discrepancy the evidence does not resolve. The Series C closed in October 2025 at $250 million, co-led by Lightspeed Venture Partners, Index Ventures, and Evantic, at a $4 billion post-money valuation. The Series D closed in July 2026 at $1.505 billion, led by Atreides Management, Index Ventures, and TCV, at a $17.5 billion post-money valuation. Total disclosed capital raised across all four rounds is approximately $1.832 billion. No pre-seed or seed round appears in the evidence; whether one existed could not be established.
The valuation trajectory is unambiguous in direction: $522–552 million in July 2024, $4 billion in October 2025, and $17.5 billion in July 2026. The jump from $4 billion to $17.5 billion occurred in approximately nine months. All three figures are investor-round valuations reported at the time of closing, not independently audited marks. The $17.5 billion figure is corroborated by Yahoo Finance, Orrick, and the company's own announcement, so it is not a single-source figure.
On revenue, several figures are on the record but they come from different sources with different levels of authority and must be read accordingly. The company's own Series D announcement, published July 15, 2026, states that Fireworks surpassed $1 billion in annualized revenue run rate at the time of the raise. A LinkedIn post by CEO Lin Qiao on the same date makes the same claim. These are founder-stated figures, not filed or audited numbers. An independent research firm estimated annualized revenue at $800 million in May 2026, up from approximately $305 million at the end of 2025; this is an analyst estimate, not a primary filing. The Wall Street Journal reported, at the time of the Series C in October 2025, that annual recurring revenue had surpassed $280 million. A July 2025 report cited annualized revenue of more than $200 million. No filed revenue figure, no audited financial statement, and no profitability disclosure appears anywhere in the evidence. The company is private and files no public accounts.
That analyst characterizes this as approximately 4x growth over the period. No gross margin, EBITDA, or cash-burn figure has been put on the record by anyone.
The gap between the Series B (July 2024) and Series C (October 2025) is fifteen months. The gap between the Series C and Series D (July 2026) is nine months. It is equally consistent with a company raising opportunistically while investor appetite is strong. The revenue trajectory, if the founder-stated figures are taken at face value, points toward the opportunistic-raise interpretation rather than distress, but no independent verification of those figures exists in this record, and a reader acting on them should treat them as unaudited management claims.
Who Pays Them, And Who They Are Up Against
Named customers, apparent concentration, and the competitive field as the record shows it.
Named enterprise customers appearing in Fireworks AI's own October 2025 Series C announcement include Samsung, Uber, DoorDash, Notion, Shopify, and Upwork — all cited in company-authored material from that date, which is a claim about relationships that existed when the post was written, not a verified current contract. The customer page references speculative decoding powering faster code edits and cites at least one case study, though the evidence does not supply the customer name or date for that specific study. An outside usage tracker dated August 2026 detected 368 companies using Fireworks AI infrastructure, suggesting the named logos represent a fraction of a broader base. The Series C announcement also states the customer base grew 10x between the Series B (July 2024) and October 2025, reaching over 10,000 companies.
That gap matters for an acquirer: if a handful of the named enterprise accounts drive the bulk of revenue, losing even one post-acquisition could materially impair the business. This could not be established from the evidence available as of 2026-09-08.
On sector mix, the named customers span e-commerce (Shopify, DoorDash), productivity and collaboration (Notion, Upwork), and consumer electronics (Samsung), with Uber representing ride-sharing and logistics. The HumanX speaker profile also cites Uber as a named enterprise user. No single vertical dominates the named list, though the evidence does not supply a breakdown of revenue by sector.
In competitive positioning, Fireworks AI is consistently compared against Together AI, Groq, Baseten, and general cloud inference offerings from AWS, Azure, and Google. A January 2026 Reddit thread from production users describes Fireworks as competitive on price-performance but includes at least one user calling the company "shady" — unverified, single source, and self-selected. A DigitalOcean comparison piece from August 2026 frames Fireworks as a specialist inference provider rather than a full-stack platform, noting the absence of broader developer tooling as a differentiator for alternatives. The Lightspeed investment memo from October 2025 positions Fireworks as aiming to power every generative AI application, which is investor framing rather than an independent assessment.
Review-site signals are thin and mixed. Trustpilot shows a 2.3 score from 8 reviews, too small a sample to draw conclusions. G2 feedback flags a difficult learning curve and weak onboarding. A December 2025 Reddit post describes unresolved critical errors in the speech-to-text product persisting for over a month — a specific, dated complaint about a non-core product line that the evidence does not show was subsequently resolved. For Nvidia, the relevant question is whether enterprise customers at the scale of Uber and Shopify are sticky enough to survive an ownership change, and the evidence does not answer that.
Warning Bells
What is actually wrong or worth worrying about here, what the evidence for it is, and what it does to this specific decision.
The pace of that growth is itself the first thing to interrogate, because it compresses the time available to verify whether the underlying business is as durable as the headline numbers suggest.
WARNINGS — STRONGEST FIRST
The revenue figures rest almost entirely on the company's own announcements and a single analyst estimate, with no audited financials in evidence. Fireworks AI's Series D blog post dated July 15, 2026 claims the company "surpasses $1 billion in annualized revenue run rate." A single secondary analyst source estimated $800 million in annualized revenue as of May 2026, up from approximately $305 million at end-2025. The Wall Street Journal reported over $280 million in annual recurring revenue at the time of the Series C in October 2025. These figures are consistent in direction but none is independently audited. For Nvidia, which would be acquiring the company outright, the absence of verified financials means the $17.5 billion acquisition price — implied by the Series D valuation reported by Yahoo Finance on July 16, 2026 — cannot be stress-tested against confirmed revenue, margin, or churn data from the evidence available here. This is a single-source / unaudited-figures concern on the most material number in the deal.
The business is structurally dependent on GPU supply that Nvidia itself controls, which creates a circular and potentially fragile cost base. Fireworks AI's product is inference speed and throughput on open-weight models; its competitive differentiation — including the custom FireAttention kernel cited in a Reddit thread from July 2026 as delivering 80% more throughput on AMD MI300s than Nvidia's NIM — means the company has already begun optimizing for non-Nvidia hardware. Acquiring Fireworks AI would bring Nvidia a platform whose engineering team has demonstrated it can route workloads away from Nvidia silicon. Whether that capability is an asset (Nvidia absorbs it and neutralizes a competitive threat) or a liability (the team's incentive to optimize for AMD disappears post-acquisition, alienating customers who chose Fireworks partly for hardware-agnostic performance) is a strategic question the evidence does not resolve. This is an inference from a documented technical pattern; confidence is moderate.
The company's infrastructure runs across multiple cloud providers — AWS, Google Cloud, Oracle, and Microsoft Azure — rather than on owned hardware. The AWS case study page and the Oracle customer page both confirm Fireworks AI as a customer of those platforms. The Microsoft Azure announcement from March 11, 2026 confirms a Foundry integration. On the day Nvidia owns this company, it inherits significant cloud spend obligations to competitors. The magnitude of that spend is not disclosed in any evidence available here and could not be established as of 2026-09-08.
Reliability complaints from developers are documented and unresolved in at least one case. A Reddit thread from December 2025 describes critical, confirmed errors in Fireworks AI's speech-to-text service that remained unfixed for over a month. The Trustpilot page for fireworks.ai shows a 2.3 score from 8 reviews. Third-party uptime trackers report a 30-day uptime of 99.67% as of the current period, per PulsAPI, which implies meaningful downtime at the scale Fireworks now operates. These are corroborated across independent sources. For Nvidia, which would be absorbing Fireworks AI's enterprise customer relationships — including Samsung, Uber, DoorDash, Notion, Shopify, and Upwork, named in the Series C announcement — any reliability gap becomes a reputational risk attached to the Nvidia brand.
The company is small relative to its valuation and the complexity of the acquisition. One private-market company profile lists 148 employees, and LinkedIn describes the company as 51–200 employees. The Sequoia job board and LinkedIn jobs page show active hiring across EMEA, GTM, security, and HR functions, indicating the operational infrastructure is still being built. A company at $1 billion ARR with under 200 employees and no dedicated security operations lead until now — ZipRecruiter lists this as a first hire — carries integration risk that is disproportionate to its revenue scale. Corroborated across multiple job-board sources.
The founding team's concentration risk is real. Lin Qiao, CEO and co-founder, is the dominant public face of the company, previously Head of PyTorch at Meta. Chenyu Zhao is listed as co-founding CTO. No evidence of a deep second-tier leadership bench was found in the records searched. SVP Finance Steven Wu joined in February 2026, meaning the finance function is less than seven months old at a company now claiming $1 billion ARR. Single-source on the finance hire; the founding concentration is corroborated.
No formal legal proceedings, regulatory enforcement actions, or sanctions matches were returned in the court, legal, regulator and enforcement records searched, or the global LEI index. No entry for Fireworks AI, Inc. was found on the UK company register. The absence of findings in those record sets is a statement about what was searched, not a clearance.
THINGS THAT LOOK LIKE WARNINGS BUT ARE NOT — FOR THIS READER
The AMD MI300 optimization work, noted above as a structural concern, is simultaneously an asset for Nvidia to assess rather than a pure liability. A platform that has proven it can extract superior throughput from competing silicon — 80% more throughput and 60% faster latency than Nvidia's NIM, per the Reddit thread — has engineering capability that Nvidia could redirect toward its own hardware stack. Whether that team would stay and redirect is a retention question, not a technology question.
The concern is verification, not the step-up itself.
The multi-cloud infrastructure dependency, while a cost and competitive complexity issue, is not evidence of operational dysfunction. It is the standard architecture for an inference platform at this stage and would be rationalized post-acquisition in any integration plan.
Corporate Standing
Whether the company legally exists, is in good standing, and is filing what it must.
Fireworks AI, Inc. is the legal entity name appearing consistently across funding announcements, investor communications, and the company's own published materials, including the Series D announcement and Orrick's deal confirmation. The company is headquartered in San Mateo, California, with Redwood City also cited in earlier coverage. It was founded in 2022, per a private-market company profile.
No entry for Fireworks AI was returned by the global LEI index or the UK company register. Those are absences in those specific registers over their current periods — they do not establish that no legal entity exists, and they do not confirm good standing. The LEI absence means no Legal Entity Identifier has been registered or published under this name, which is unremarkable for a private US company that has not issued debt or securities requiring LEI registration. The UK-register absence confirms only that no UK-registered entity under this name was found; the company's operations appear to be US-domiciled.
No company register search of a US jurisdiction — Delaware, California, or otherwise — was conducted for this report. For a company of this size, the operative question is Delaware incorporation and good standing, since virtually all US venture-backed companies of this profile incorporate there. That certificate of good standing, the Delaware entity number, and the registered agent on file were not obtained and could not be established from the evidence available. This is a material gap for an acquisition: Nvidia's counsel will need a certified certificate of good standing from the Delaware Secretary of State and confirmation that the entity has not been administratively dissolved or had its charter voided for non-payment of franchise tax.
Pre-IPO secondary market listings on Nasdaq Private Market and Forge Global confirm that Fireworks AI stock is being traded in secondary markets, which implies the existence of a recognized legal entity with issued equity — but secondary market activity is not a substitute for a primary register search.
No regulatory enforcement actions, sanctions matches, or exclusion notices were found in regulator and enforcement records. That is an absence in those records, not a clearance.
The specific documents needed before closing are: a certified Delaware certificate of good standing, the current certificate of incorporation and all amendments, confirmation of the registered agent, and a cap table certified by the company's counsel reflecting the post-Series-D structure. None of those were available in the evidence supplied.
Who Controls It
Registered owners and persons with significant control.
Fireworks AI, Inc. is incorporated in the United States — the company's own materials consistently use that legal name and the entity appears on secondary market platforms including Nasdaq Private Market and Forge Global, confirming it exists as a private U.S. corporation. No entry was returned from the UK company register, consistent with the company having no UK registration. No LEI filing was located. No U.S. state of incorporation, no registered agent name, and no cap table have been disclosed in any evidence available as of 2026-09-08.
Beneficial ownership is the central gap for an acquirer. The evidence identifies the investor syndicate but not the percentage stakes each holds. Sequoia Capital led the Series B at a $522 million valuation in July 2024. The Series C in October 2025 — $250 million at a $4 billion valuation — was co-led by Lightspeed Venture Partners, Index Ventures, and Evantic. The Series D in July 2026 — $1.505 billion at a $17.5 billion valuation — was led by Atreides Management, Index Ventures, and TCV. No source discloses the resulting ownership percentages for any of these investors, nor the aggregate dilution experienced by founders across four rounds.
Co-founder and CEO Lin Qiao is the most prominent named individual. Co-founder Benny Chen and co-founder Chenyu Zhao (described in one source as founding CTO) are also identified on the team page. The founding team's retained equity after a Series D at $17.5 billion — following rounds at $522 million, $4 billion, and now $17.5 billion — is not disclosed. Given the scale of dilution across four institutional rounds totalling well over $1.8 billion raised, it is a realistic possibility (40–50%) that no single founder retains a majority stake, though this cannot be confirmed without the cap table (Confidence: Low — no ownership data in evidence).
For Nvidia, the practical consequence is that a signed acquisition requires consent or tender from multiple institutional investors across at least four rounds, each with its own liquidation preferences, pro-rata rights, and potentially board seats. The specific governance rights — board composition, protective provisions, drag-along and tag-along terms — are not in any evidence available here.
What would answer this: the Delaware certificate of incorporation and any amendments, the most recent stockholders' agreement, the cap table as of the Series D close, and board minutes or a governance summary from Fireworks AI's counsel. These are standard items in a data room and should be the first documents requested before any offer is tabled.
How Complicated The Cap Table Is
Preference stacks, secondary sales and option overhang — the things that decide who actually gets paid at close.
Fireworks AI, Inc. is a Delaware-incorporated private company that has completed four priced equity rounds in roughly 28 months, each at a materially higher valuation than the last. The Series A closed at an undisclosed valuation in March 2024; the Series B closed in July 2024 at a $522 million post-money valuation; the Series C closed in October 2025 at $4 billion; and the Series D closed in July 2026 at $17.5 billion on $1.505 billion raised. Total disclosed capital raised across all rounds exceeds $1.83 billion. Each of those rounds almost certainly carries a liquidation preference — standard for venture-backed Delaware corporations — meaning that at acquisition close, preferred shareholders collect their preference stack before any proceeds reach common stockholders or option holders. The aggregate preference stack is not disclosed in any evidence available here.
The Series D was led by Atreides Management, Index Ventures, and TCV, with Index Ventures also having co-led the Series C alongside Lightspeed Venture Partners and Evantic. Sequoia Capital led the Series B. That is at least four institutional investors with board or observer rights to negotiate around, and the layering of three separate lead investors across the last two rounds alone creates a preference stack that Nvidia's deal team would need to unwind in sequence.
On option overhang: Sequoia's job board lists Fireworks AI as a 100–1,000 employee company, while LinkedIn and a private-market company profile put current headcount at roughly 148–200. A company at this growth rate and headcount, having raised at successively higher valuations, will have issued options at strike prices ranging from sub-$522 million valuation to sub-$4 billion valuation. The spread between early-grant strike prices and the $17.5 billion Series D price creates substantial in-the-money overhang — but the actual option pool size, vesting schedules, and acceleration provisions on change of control are not on the record.
Secondary sales are a specific gap. Nasdaq Private Market and Forge Global both list Fireworks AI stock for secondary trading, which means some early employees or investors have already sold shares. The volume, pricing, and whether those sales involved any right-of-first-refusal waivers from the company are unknown.
What Nvidia needs before signing: the full capitalization table with share class, liquidation preference multiples and participation rights for each series; the option pool schedule with vesting and change-of-control acceleration terms; and a complete secondary-transaction log showing any transfers that altered the preference stack or triggered co-sale rights. None of those documents are on the record.
Is The Revenue Real, And Does It Recur
Whether what this company calls revenue is contracted and repeating, or one-off work counted once.
The revenue figures in evidence are substantial and directionally consistent across multiple sources: Fireworks AI's own Series D announcement claims over $1 billion in annualized revenue run rate as of July 2026, an independent analyst estimates $800 million in annualized revenue as of May 2026, and the Wall Street Journal reported annual recurring revenue surpassing $280 million at the time of the October 2025 Series C. The trajectory is steep and the figures are corroborated by independent parties, but the nature of the revenue — whether it is contracted and repeating, or consumption-driven and revocable — is not answered by any source in evidence.
The product is a usage-based inference and fine-tuning platform. Customers pay per token on serverless workloads, or reserve dedicated GPU capacity on-demand. Billing is pre-paid credits drawn down by usage, not invoiced against multi-year commitments. That structure means the revenue figure being cited is almost certainly an annualized run-rate of consumption, not a contracted backlog. A customer running Fireworks in production today can route traffic elsewhere tomorrow without breaching a contract. Named customers including Samsung, Uber, DoorDash, Notion, Shopify, and Upwork are real enterprise relationships, but whether any of those are governed by minimum-commit agreements with meaningful termination penalties is not established in the evidence.
For Nvidia specifically, this distinction is material. Switching costs exist — fine-tuned models, custom inference pipelines, and integrations create friction — but they are not the same as a signed contract.
What would settle this: the contract schedule showing the split between committed annual contracts and pure consumption accounts, the renewal cohort data showing what percentage of prior-year revenue re-committed in the following year, and the net revenue retention rate by customer tier. None of those are public. Request them as a condition of advancing to exclusivity.
Can They Fund The Term
Funding, financial standing, and how long the money in public view lasts.
Fireworks AI has raised a total of approximately $1.83 billion across four rounds. The Series A closed at $25 million in March 2024, the Series B at $52 million in July 2024 at a $522 million valuation, the Series C at $250 million in October 2025 at a $4 billion valuation, and the Series D at $1.505 billion in July 2026 at a $17.5 billion post-money valuation, led by Atreides Management, Index Ventures, and TCV. Orrick, which advised on both the Series C and Series D, confirmed the Series D close in July 2026.
The company's own Series D announcement states that annualized revenue crossed $1 billion at the time of the raise, as of July 2026. An independent analyst estimated annualized revenue at $800 million in May 2026, up from approximately $305 million at end of 2025, implying rapid sequential growth. The Series C announcement in October 2025 cited annual revenue exceeding $280 million at that point, corroborated by the Wall Street Journal.
That capital arrived two months before this report date. Stated use of proceeds is infrastructure expansion and engineering headcount growth. Burn rate, cash balance, and the precise deployment schedule of Series D proceeds are not disclosed in any source available here; those figures would require access to management accounts or a data room.
Whether that multiple is defensible in an acquisition context depends on revenue quality and contract structure, which are addressed in the revenue section of this report.
What this section cannot establish: the actual cash balance, the burn rate against the Series D proceeds, the terms of any debt or credit facility, and whether any investor holds liquidation preferences that would affect acquisition proceeds distribution. A data room request covering the capitalization table, the Series D term sheet, and the most recent management accounts would answer those questions before any acquisition price is fixed.
If The Founder Leaves
How much of this company's ability to deliver sits with one or two named people.
Lin Qiao is the gravitational center of Fireworks AI in ways that go beyond a typical founder-CEO role. She was previously Head of PyTorch at Meta — the framework that underpins most of the open-weights model ecosystem Fireworks is built to serve — and that technical credibility is woven into how the company positions itself externally. The team page leads with her name and prior role, and the Series D announcement, the Series C announcement, and the company's public narrative around "specialized intelligence" and post-training are all delivered in her voice. The Business Insider profile from April 2026 describes her as the person running a company processing 15 trillion tokens a day; the Sequoia appearance in August 2026 frames the company's entire product thesis — post-training as the durable moat — through her argument alone.
Co-founder Benny Chen is named on the team page, and Chenyu Zhao is identified as founding CTO in a third-party directory, described as having led machine learning architecture. Steven Wu joined as SVP Finance and Business Operations in February 2026 per his LinkedIn profile. Beyond those names, the evidence does not establish a second tier of named technical or commercial leadership with independent external profiles or attributed decision-making authority.
For Nvidia specifically, the dependency question has two layers. First, Qiao's relationships with the open-weights model community — built through her PyTorch tenure at Meta — are a sourcing and partnership asset that does not automatically transfer with a corporate acquisition. Model providers, open-source contributors, and enterprise customers who chose Fireworks partly because of who runs it may reassess under new ownership. Second, the company's technical differentiation — FireAttention, speculative decoding, the inference optimization stack — was built under her and Zhao's direction. Whether that institutional knowledge is distributed enough to survive their departure is not answerable from the evidence available.
What is missing and would answer this directly: employment agreements, vesting schedules, and any retention or key-person provisions negotiated at the Series D closing. A conversation with the co-founders about post-acquisition roles, and reference calls with two or three named enterprise customers asking whether their relationship is with the platform or with the people, would resolve the practical exposure. None of those documents or conversations are in the evidence supplied.
Who Is Already Looking
Public signals that people here are moving on before you have made an offer.
The clearest public signal of potential movement is the hiring posture Fireworks itself has adopted. Open roles listed through Sequoia's job board include a Compensation Partner, an HR Business Partner for GTM, a Senior Billing Operations Analyst, and a Senior Manager for Indirect Tax — functions that a company of 51–200 people builds when it is preparing for a liquidity event or a significant ownership change, not when it is simply scaling headcount. A first Security Operations Lead is also being recruited, per ZipRecruiter, which signals that the security function has not yet been institutionalized. For Nvidia, that gap matters on day one of ownership.
The founding team's origins create a specific retention risk. Lin Qiao was previously Head of PyTorch at Meta, and the company's own team page describes the broader group as bringing "decades of AI experience from PyTorch, Meta and Google." Meta, Google, and the wider hyperscaler ecosystem are the natural re-entry points for engineers who decide an acquisition does not suit them. An April 2026 podcast description characterizes the company as "built by seven ex-" — the sentence is truncated in the evidence, so the originating employer cannot be confirmed from this source alone, but the pattern of Meta and Google provenance is corroborated across multiple sources. These are people with standing offers available to them at any time.
No named departures from the leadership team appear in any of the record sets searched — workforce trackers and review sites included. That absence is not a clearance; it reflects that private-company attrition rarely surfaces publicly until it is already consequential.
What is missing and would directly answer this question: vesting schedules and cliff dates for the founding team and senior engineers, any retention or change-of-control bonus agreements already in place, and direct conversations with the co-founders about their post-acquisition intentions. The Glassdoor profile holds only nine reviews and yields no signal on flight risk. The specific document that would resolve this is the cap table with vesting detail and any side letters attached to the Series D close in July 2026 — none of that is on the record.
Litigation and Regulatory Record
Court and regulator activity on the record.
Court, legal, regulator and enforcement records were checked for Fireworks AI. None returned litigation filings, regulatory enforcement actions, consent orders, sanctions matches, or exclusion notices naming Fireworks AI, Inc.
That absence is not a clearance. It is not a full search of federal or state court dockets. No federal docket search, no state superior court docket review, and no USPTO trademark or patent dispute search was conducted.
One item in the evidence warrants attention without being overstated. A LinkedIn post from June 2026 references a class action lawsuit alleging that a company misled investors about AI and Azure — the post appears in the context of Fireworks AI's Microsoft Foundry availability announcement, but the lawsuit described is not identified as being against Fireworks AI itself. The post does not name Fireworks AI as a defendant, and attributing that litigation to Fireworks AI would be an inference the evidence does not support. Identity of the defendant in that action could not be established from the evidence provided.
The regulatory exposure most relevant to Nvidia's acquisition decision — and not answered by the evidence — is whether Fireworks AI's commercial fine-tuning service has received any inquiry from the FTC, the Copyright Office, or any model licensor regarding the terms under which customer data is used to fine-tune third-party open-weights models. That question is material because the fine-tuning product is a core revenue driver, and any restriction on its operation post-close would directly impair what Nvidia is buying. The specific documents needed are: any demand letters or regulatory correspondence in Fireworks AI's files, a representation from counsel that no such inquiries are open, and a review of the indemnification terms in Fireworks AI's customer agreements covering IP claims arising from fine-tuned outputs.
What Breaks On Close
Change-of-control clauses, key contracts and licences that do not survive a change of owner.
The evidence available does not contain Fireworks AI's customer contracts, partner agreements, or any licence terms that would reveal change-of-control consent requirements. What follows identifies the specific exposure points that a Nvidia acquisition would trigger, and what remains unresolved.
The most material risk sits in the Microsoft Foundry integration announced in March 2026 and the Oracle OCI partnership and AWS relationship. Fireworks distributes inference capacity across all three of Nvidia's largest cloud competitors. Each of those arrangements almost certainly carries change-of-control provisions — standard in cloud partnership agreements — that would give Microsoft, Oracle, and AWS the right to renegotiate or terminate on a transfer of ownership to Nvidia. Whether those clauses are consent-based, termination-based, or merely notification-based cannot be established from the evidence supplied. The specific partnership agreements have not been produced and were not reachable through the record sets searched.
The named enterprise customers — Samsung, Uber, DoorDash, Notion, Shopify, and Upwork — represent a customer base that, at $1 billion ARR, is generating material revenue under contracts whose change-of-control terms are unknown. Enterprise SaaS agreements routinely require customer consent to assignment. If even a fraction of the top-decile customers hold consent rights, the revenue figure that justifies the acquisition price becomes contingent on post-close negotiations Nvidia does not control.
The open-source model licensing structure presents a separate question. Fireworks serves open-weights models under licences granted to Fireworks AI, Inc. as the operating entity. Some of those licences, particularly Meta's commercial model licence, contain restrictions on use by entities above certain revenue or user thresholds, and some versions have contained provisions specific to the licensee. Whether the licences survive a change of beneficial ownership to Nvidia, or whether Nvidia's own scale triggers a different licence tier or restriction, requires direct review of each model licence against the post-acquisition entity structure.
The three documents that would answer this section are: the full text of the Microsoft Foundry, Oracle OCI, and AWS partnership agreements; the standard master service agreement Fireworks uses with enterprise customers, specifically the assignment and change-of-control clause; and the model licence agreements for every open-weights model currently in production serving. None of those were available in the record sets searched.
What Happens To You If They Are Bought
Whether this company looks like an acquisition target, and what a change of control would do to your price, your data and your contract.
Fireworks AI is the acquirer in this scenario, not the target — but the question is what happens to Nvidia's position as a customer, partner, or prospective buyer if someone else acquires Fireworks first, or what a Nvidia acquisition of Fireworks would do to Fireworks' existing relationships and contracts.
The evidence establishes that Fireworks AI, Inc. is a private Delaware-incorporated entity (the "Inc." designation appears consistently across its own blog and legal counsel announcements, including Orrick's Series D closing notice). Its cap table now includes Atreides Management, Index Ventures, TCV, Lightspeed Venture Partners, Sequoia Capital, and Evantic, following the July 2026 Series D at a $17.5 billion post-money valuation. That investor syndicate is material: a change-of-control transaction at a price that does not satisfy those investors' return expectations would face internal resistance. The valuation step from $4 billion in October 2025 to $17.5 billion in July 2026 compresses the window in which an acquirer can offer a price that clears the syndicate without paying a substantial premium to the already-elevated Series D entry point.
The customer base is the most immediate change-of-control risk. Named customers include Samsung, Uber, DoorDash, Notion, Shopify, and Upwork, and the platform is described as serving over 10,000 companies as of October 2025. Fireworks competes directly against cloud-native inference offerings from Google, Microsoft, and AWS — all of whom are simultaneously infrastructure partners. Microsoft Azure lists Fireworks as a generally available partner on Microsoft Foundry, and Google Cloud has published a case study on the partnership. An Nvidia acquisition would place those cloud partnerships under direct competitive pressure: Microsoft, Google, and AWS would have a strong commercial incentive to migrate Fireworks-dependent workloads to their own inference stacks the moment Nvidia ownership became public. The contracts governing those partnerships — their change-of-control clauses, termination rights, and exclusivity provisions — are not in the evidence and could not be established from the record sets searched.
The specific documents needed to answer this question fully are: the investor rights agreement and any drag-along or co-sale provisions from the Series D closing, the enterprise customer master service agreements (particularly change-of-control and termination-for-convenience clauses), and the partnership agreements with Microsoft Azure, Google Cloud, and AWS. None of those were available in the record sets searched for this report.
What Has Happened Lately
Dated public events, most recent first.
August 2026 brought the most strategically significant public signal since the Series D: Lin Qiao used a Sequoia Capital platform appearance on August 12–13 to argue that post-training — not prompting — is the durable competitive layer for AI application companies, and that the cost of replicating software from a screenshot has collapsed to near-zero. The argument is a direct product positioning claim: Fireworks intends to be the platform on which customers bake proprietary intelligence into open models, not merely a commodity token-serving layer. For Nvidia, this matters because it signals where Qiao believes the moat sits — and therefore what you would be buying versus what you would be inheriting as a commodity business.
In July 2026, Fireworks closed a $1.505 billion Series D at a $17.5 billion valuation, led by Atreides Management, Index Ventures, and TCV. The company simultaneously announced it had crossed $1 billion in annualized revenue run rate and was serving more than 40 trillion tokens per day. An independent analyst estimated annualized revenue at approximately $800 million as of May 2026, up from roughly $305 million at end of 2025, which implies the $1 billion ARR figure was reached sometime between May and the July announcement. The valuation moved from $4 billion at the October 2025 Series C to $17.5 billion in nine months — a 4.4× step-up that reflects the revenue trajectory but also embeds significant growth expectations into any acquisition price.
In March 2026, Fireworks launched on Microsoft Azure via the Foundry program, making its inference available as a managed service inside Azure. That distribution agreement is a material commercial relationship that Nvidia would need to assess for change-of-control provisions on day one of ownership.
Also in April 2026, Business Insider reported the company was processing 15 trillion tokens per day at that point — a figure that had grown to 40 trillion by the July Series D announcement, indicating roughly 2.7× volume growth in under four months.
No layoff events, enforcement actions, or leadership departures appear in the record sets searched for the period covered by this report. What is not established from available evidence is the specific contractual structure of the Microsoft Foundry arrangement, the revenue concentration among named customers such as Samsung, Uber, DoorDash, Notion, Shopify, and Upwork, and whether any investor rights attached to the Series D include acquisition consent or drag-along provisions that would affect deal mechanics.
Sources
Every source this report rests on, so any sentence in it can be checked.
Every source cited above, deduplicated and grouped. Each was read as it stood on 2026-09-08; a page can change after that date.
fireworks.ai
linkedin.com
reddit.com
docs.fireworks.ai
aws.amazon.com
wsj.com
businesswire.com
youtube.com
orrick.com
computeprices.com
pricepertoken.com
businessinsider.com
siliconangle.com
finance.yahoo.com
datacenterdynamics.com
humanx.co
digitalocean.com
lsvp.com
trustpilot.com
g2.com
oracle.com
azure.microsoft.com
pulsapi.com
jobs.sequoiacap.com
ziprecruiter.com
nasdaqprivatemarket.com
forgeglobal.com
podcasts.apple.com
glassdoor.com
cloud.google.com
41 distinct sources across 30 domains.