Ramp's October 2026 AI category data has Fireworks AI as the most-purchased Model Serving and Inference vendor. 27% of buyers in the category. It is also the most fragmented category on the list, so 27% is a lead, not a lock.
Then I opened the ads. Of 231 live LinkedIn ads, 23 sell inference or free credits. 155 sell something else: training your own model, a conference called Forge, and customers who trained one.
The inference winner advertises training. Search is where the inference ads went.
Fireworks AI ads: key findings
- 231 live LinkedIn ads, 108 distinct headlines, 22 of them video
- 115 Google ads, 112 text and 3 video, shown in 62 countries
- 0 Meta ads
- 155 of 231 LinkedIn ads (67%) are about training, owning a model, Forge or customers who trained one
- 23 of 231 (10%) are inference-first: model launches and free credits
- 34 ads for Forge, the 3 November conference, and 36 for the Nexus cost-control product
- 48 of 115 Google ads (42%) are model launches, from DeepSeek to GLM to Kimi
- 12 Google comparison ads: Baseten, Groq, Together, Cerebras and Kimi K3 against a closed model
- 68 Google ads were still showing in October 2026
LinkedIn is where Fireworks sells the next product
Almost none of the 231 LinkedIn ads explain what inference is. The buyer already has a model. The ads tell them they should not be renting it.
Eight messages, mutually exclusive
| Message cluster | Ads | Share | Headline examples |
|---|---|---|---|
| Train and own your model | 70 | 30% | "Training API + Fireworks Lab are now generally available." (18), "Build intelligence that's uniquely yours." (18), "Make the Model Your Edge" (6) |
| Customer case studies | 46 | 20% | "Build your own legal frontier." (5), Cursor's Composer 2, Genspark, Heidi Health |
| Events | 39 | 17% | "Forge: Build Your Own Frontier" (24), "Forge | November 3 | San Francisco" (8), First Pour in Paris |
| Cost control | 36 | 16% | "Spend and ship, or ration tokens and stall?" (4), "More control over AI coding spend" (4), "Same workflow. 54% less AI spend." (2) |
| Model launches and benchmarks | 15 | 6% | "Kimi K3 is Live: see how it compares to Opus 5" (3), Muse Glimmer, Qwen, DeepSeek |
| Founder and thought-leader posts | 9 | 4% | Lin Qiao, the CEO, on rented versus owned intelligence |
| Free credits and generic inference | 8 | 3% | "Start Building with $100 of Free Credits" (2) |
| Jobs, follow, message and document formats | 8 | 3% | Four job ads, two message ads |
The most-purchased inference vendor spends 3% of its LinkedIn ads on "start building with free credits." The old acquisition message is the smallest bucket on the page.
The copy: you write the loop, we run the rest
The Training API ad says "Frontier-level training shouldn't require frontier-lab infrastructure. You write the training loop and reward. Fireworks handles the trainer, rollouts, and infrastructure underneath it."
That is the same company that sells inference. The new product is a reason to stay, not a reason to leave. Train on Fireworks, serve on Fireworks.
Harvey, Cursor, Genspark, Heidi
46 case-study ads, and they are not about speed. They are about a customer who built a model.
| Customer | Ads mentioning them | The claim |
|---|---|---|
| Cursor | 15 | Composer 2, trained with large-scale RL, "6-10x lower inference cost" |
| Harvey | 13 | "Nearly twice the tasks completed at effectively flat cost" |
| Heidi Health | 13 | Latency cut 3.5x, from 25 seconds to 7, proof of concept to production in 4 weeks |
| Genspark | 10 | Gen-1 Slides, "About 90% less per deck than Opus 5" |
Counts overlap because one ad can name Cursor and a conference. The pattern does not. Every named customer is a company whose own model is the product.
The cost ads are a second product
36 ads cover cost control, and 27 name Nexus, a product that routes coding-tool traffic from tools like Claude Code and Codex to open models. "Spend and ship, or ration tokens and stall?" is the question. "Frontier intelligence. Half the bill." is the answer.
The buyer here is a head of engineering with a token budget, not a developer with a prompt. A separate Fireworks ad quotes Ramp Economics Lab: four of the top 10 trending vendors are AI inference platforms, because teams are getting deliberate about which model handles which workload.
The CEO is an ad unit
Nine posts run as founder or thought-leader content. Four run under Lin Qiao's own name, two of them Article ads, and two video ads carry her line that the future "will not be a small number of AGI models dominating the world... it will be millions of specialized" ones.
One Article ad argues that a company that owns its intelligence gets better every week and one that rents it is a policy change from starting over. That is the whole LinkedIn campaign in two sentences.
Google is where the inference ads went
The 115 Google ads are almost all text. 112 text, 3 video. First shown 28 January 2025, and 68 were still showing in October 2026.
| Headline cluster | Ads | Share | What it says |
|---|---|---|---|
| Model launches and model pages | 48 | 42% | "DeepSeek V4 has Landed," "GLM 5.3 is open-weight and live on Fireworks," "Kimi K2.6 is on Fireworks AI," "Ember-1: Only on Fireworks" |
| Own and train your model | 20 | 17% | "Own your own AI with Fireworks," "Beat the Frontier on Your Domain," "Learn About Training with Fireworks" |
| Generic open-source inference | 15 | 13% | "Open-source AI Inference - Always Open. Always Fast" |
| Brand name only | 13 | 11% | "Fireworks AI" (10 near-identical) |
| Head-to-head comparisons | 12 | 10% | "Fireworks vs. Baseten," "vs. Groq," "vs. Together," "vs. Cerebras," "Compare Kimi K3 vs. Fable 5" |
| Case studies and cost proof | 7 | 6% | "How Juicebox Cut Latency by 80%," "Frontier Intelligence. Half the Bill." |
Open-model releases are search events
48 ads, one every time a lab ships. Kimi appears in 21 ads, DeepSeek in 13, GLM in 11, Qwen in 5. Many link to a model-specific page, like a DeepSeek V4 start page or a GLM 5.3 serverless page.
The timing is the tell. "Kimi K3 is Coming Soon" and waitlist ads ran 20 to 27 July. The Kimi K3 comparison ads first showed on 27 July. Fireworks was in the results before the model was live and stayed there after.
The comparison layer
"Fireworks vs. Baseten" shows up in four separate ads. Groq, Together and Cerebras each get one or two. These people are comparing vendors and already typing a rival's name. Four more ads compare Kimi K3 to a closed model.
The strange part is how many of those pages say "Own." 20 headlines carry "Own your own AI" or training language, even on a query like "open-source AI inference." The LinkedIn story leaks into search.
Why the split makes sense
| Ads | 231 | 115 |
| Train and own | 70 (30%) | 20 (17%) |
| Model launches | 15 (6%) | 48 (42%) |
| Comparison to rivals | 0 | 12 (10%) |
| Events | 39 (17%) | 0 |
| Cost control | 36 (16%) | 7 (6%) |
Google buys intent. Someone types "DeepSeek V4 API" and Fireworks answers with a page for that model. LinkedIn buys a conversation with a team that has already picked a serving vendor, so the message is the next purchase.
Meta is not in the plan. No Meta ads at all.
What actually makes this work
Fireworks does not market what it won. It markets what it wants to sell next.
The 27% of buyers already have the serving relationship. The training, the Nexus routing and the conference all build on it. Each ad gives that buyer a reason to put one more workload on the same vendor.
Search runs the reverse job. Every new open model is a fresh query, and Fireworks is the first ad on it. The two channels do not repeat each other, and neither one runs the "start building" message the category normally runs.
If your category's winner is a utility, look at what it sells on its second channel.
Break down your own market
The live ads are in the LinkedIn ads library and the Google ads library, and every other advertiser is in the Sprites ads library. Creative analytics does the same for your own account and competitors: every live ad, the hooks, and what changed this month.
More from the AI category winners: the Viktor AI ads teardown, the Wispr Flow ads teardown and the Granola ads teardown. Ready to run your own? Start here.