Fireworks Needs 400 Engineers by December. The US vLLM Pool Is 14.

On July 16, 2026, Fireworks AI closed a $1.505B Series D at a $17.5B valuation and told the press it would triple its 200-person team before December. That is roughly 400 net hires in under six months, aimed at a talent pool that any honest sourcer can count on their fingers. Together AI ($800M, July 2) and Baseten (also tripling in 2026) are hunting the same names.

The math on Fireworks AI hiring does not close

Fireworks needs about 400 engineers in five months, but the US pool of people who have actually shipped production LLM serving code is under 200. The rest of the plan is either offshore, adjacent, or aspirational.

Here is the supply-and-demand picture in one table, pulled from the research and from Refolk's index of professional profiles:

Fireworks by itself is oversubscribed roughly 4x against the narrowest tier. Add Together, Baseten, and Inferact (the January 2026 vLLM commercialization spinout that raised $150M seed) and combined 2026 plans point to somewhere between 800 and 1,200 inference-adjacent hires against a US pool that almost certainly sits under 500 people who have merged perf-critical PRs.

14

Why the ex-Meta PyTorch well is already dry

The ex-Meta PyTorch pipeline that made Fireworks possible is essentially exhausted. Lin Qiao founded Fireworks in 2022 with six co-founders, all former Meta engineers, and the team has been recruiting ex-PyTorch peers for four straight years.

The mechanism is not mysterious. A static internal talent pool cannot survive four years of continuous poaching by four well-funded startups (Fireworks, Together, Baseten, Modal) plus NVIDIA and every hyperscaler with a serving team. The Meta-resident PyTorch engineers who wanted to leave have already left. The remaining ones stayed for a reason: comp, immigration status, or a project they will not walk away from.

That reshapes where the marginal hire in Q4 comes from:

• Red Hat, which now anchors the vLLM contributor base after the IBM acquisition
• xAI, Skywork, Oracle, and LinkedIn, which supply SGLang's core contributors
• Berkeley Sky Computing Lab / LMSYS, where both Woosuk Kwon (vLLM) and Lianmin Zheng (SGLang) trained under Ion Stoica
• NVIDIA's TensorRT-LLM team, structurally the hardest to poach because of RSU cliffs
• Inferact's early hires, who will be reachable in ~18 months when the seed vesting cliff hits

If your sourcing playbook this fall still leads with "ex-Meta PyTorch," you are searching the same LinkedIn slice that Fireworks recruiters cleared out in 2023.

The real bottleneck is reviewers, not contributors

The scarce resource is not people who write CUDA kernels. It is people who are trusted to merge them. vLLM issues typically get a maintainer response within 12 hours to 3 days; SGLang runs 3 to 5 days. Behind those numbers sit single-digit head counts of gatekeeper maintainers per project.

Hiring one of those maintainers is worth roughly ten mid-level "CUDA-curious" engineers, for three reasons:

1. They already know the codebase's failure modes, so they onboard in weeks, not quarters.
2. Their PR review authority converts directly into your team's throughput.
3. They carry contributor networks. A maintainer who leaves Red Hat for Baseten pulls two or three regular contributors within six months.

The named cross-project contributors are the shortest path to this tier:

• comaniac (OpenAI): 17 early PRs to SGLang plus 77 PRs total to vLLM
• ShangmingCai (Alibaba Cloud): 18 PRs to vLLM, then 52 PRs to SGLang
• CatherineSue (Oracle): 4 bug-fix PRs to vLLM, then 76 PRs to SGLang as a core contributor

None of these people will show up in a LinkedIn Boolean built around "vLLM" as a keyword. That search returns 125,680 profiles globally, and most of them are marketers, PMs, and infra-adjacent ICs who dropped the term into a headline once. GitHub cross-repo activity beats LinkedIn keyword search roughly 10:1 for this specific role, which is the exact gap Refolk closes: describe the person in plain English (say, "US engineer who merged perf PRs to both vLLM and SGLang in the last 18 months") and get a ranked shortlist instead of a keyword sludge.

China is half the pool and mostly un-hireable at scale

Roughly 52% of SGLang contributors and 33% of vLLM contributors are based in China, which means the US-visa-eligible slice of the true expert pool is close to half of the headline number. If 194 developers have contributed to both vLLM and SGLang, the US-resident dual-stack pool is likely 80 to 115 people.

That number is small enough to name. It is also small enough that any US-only sourcing strategy will lose. The companies that survive the fall are the ones that stand up a hiring surface where the other half of the pool actually lives:

• Singapore or Vancouver hubs to hire from mainland China without H-1B lottery risk
• EU nodes (Berlin, London, Zurich) for Skywork and Alibaba Cloud diaspora engineers who want a working-hours change
• Remote contractor arrangements with the top individual contributors, followed by relocation offers at 12 months

Fireworks has not publicly committed to non-US hubs at the scale of 400 hires. Together's "50x infra footprint" language suggests it will. Baseten is quieter. The recruiter who ships a compliant Singapore entity in Q3 will out-hire the recruiter with the biggest US comp band.

The under-recruited feeder employers

The passive candidates most recruiters skip look nothing like classic FAANG ML sourcing targets. Refolk's index of the 14 US PyTorch-plus-inference profiles shows the top employers are Amazon (2), Salesforce, DigitalOcean, Oracle, and Akamai, concentrated in NYC (2), the Bay Area, and Austin.

That is a strange list only if you expect a Meta / Google / Anthropic distribution. It makes sense the moment you accept that production LLM serving at scale now happens inside enterprise clouds:

• Oracle Cloud Infrastructure hosts SGLang core contributors and runs meaningful inference workloads for its own customer base
• DigitalOcean's GPU droplet effort has quietly hired serving engineers who ship vLLM patches
• Akamai (post-Linode) has an edge-inference team that reads exactly like Fireworks' target profile
• LinkedIn's kernel and inference teams are named as SGLang core contributor employers
• Snowflake and Databricks serving teams sit in the same skill window and are rarely counted

None of these people call themselves "LLM infrastructure engineers" in their headlines. They call themselves "Principal SWE, Cloud AI" or "Staff Engineer, Platform." Boolean fails here. This is the second place Refolk earns its keep: you can ask for "senior ICs at enterprise clouds who ship to open-source inference stacks" and get the twelve names that matter, not the twelve thousand that don't.

The ex-Meta PyTorch well is dry. The next tier lives at Red Hat, Oracle, DigitalOcean, and Berkeley, not Menlo Park.

What Fireworks, Together, and Baseten are actually competing on

The three firms are chasing the same 100 people and the same 20 maintainers, which means the hiring competition will be decided on non-comp levers. Cash is roughly equalized. Fireworks passed $1B ARR (up 5x YoY) and is serving 40+ trillion tokens daily. Baseten tripled annualized revenue in one quarter at up to a $13B valuation. Together sits at $8.3B post-money. All three can pay.

The differentiators the top-tier candidates actually weight, based on the customer and contributor signals in the research:

Candidates in this tier care about which customers hit their serving path in production. The Harvey and Cursor overlap between Fireworks and Baseten is not accidental; it is the exact application layer that generates the interesting perf problems. Recruiters who lead outreach with "your PR would land in Cursor's autocomplete latency budget" close faster than recruiters who lead with base salary.

A concrete 60-day sourcing plan

The next 60 days determine who staffs Q1 2027, because offers extended in September clear diligence and start in January. Here is the plan that actually maps to the pool sizes above.

1. Map the 194 dual-stack contributors by GitHub handle, then resolve to real names and current employers. Skip anyone already at Fireworks, Together, Baseten, Inferact, or Modal.
2. Filter to the ~80 to 115 US-resident and US-visa-eligible names. This is your total addressable pool for on-shore hires.
3. Rank by merge authority, not commit count. Cross-reference maintainer lists on vLLM and SGLang. The top 20 here are worth the top 200 elsewhere.
4. Build a parallel list for the enterprise-cloud feeders: Red Hat, Oracle, LinkedIn, xAI, DigitalOcean, Akamai, Snowflake. Use plain-English search rather than Boolean; the headlines will not match.
5. Stand up a Singapore or Vancouver entity to hire from the ~half of the pool based in China. Do this before you extend the first offer in the US pool, because your comp bands need to be internally consistent.
6. Reach out with a specific PR or benchmark, not a job description. These engineers are 20-plus recruiter messages deep this quarter.

The core sourcing motion in steps 1 through 4 is exactly what Refolk was built for: describe the engineer you want in one sentence, get a ranked list across GitHub, LinkedIn, and the open web, and skip the two weeks of Boolean tuning that keyword tools force. AI infrastructure recruiting has never rewarded generalist sourcing, and this quarter it will punish it.

FAQ

How many engineers does Fireworks AI actually need to hire by December 2026?

Roughly 400 net new engineers. CEO Lin Qiao told the press on the July 16, 2026 Series D announcement that the company plans to triple its 200-person team before year-end. Fireworks did not publish a role-by-role breakdown, but given the company's product (an inference cloud that grew from 15 trillion to over 40 trillion daily tokens), the majority of those hires are inference, kernel, and serving engineers rather than GTM.

Is "vLLM contributor" a useful sourcing filter on LinkedIn?

No, not by itself. The keyword "vLLM" returns 125,680 profiles globally, dominated by people who mentioned the term once in a headline or a project bullet. The useful filter is GitHub merge activity on vllm-project/vllm or sgl-project/sglang, cross-referenced with current employment and location. Only about 194 developers globally have contributed to both projects, and roughly half of that dual-stack pool is US-based.

Where should I source LLM inference engineers if I'm not Fireworks, Together, or Baseten?

Look at the feeder employers the big three under-index on: Red Hat (vLLM backbone), Oracle Cloud, LinkedIn's kernel and inference teams, xAI, Skywork, DigitalOcean, and Akamai. Add the Berkeley Sky Computing Lab and LMSYS alumni network, which is the single densest node because Woosuk Kwon (vLLM) and Lianmin Zheng (SGLang) both trained there under Ion Stoica. These candidates rarely self-describe as "LLM infrastructure engineers," so keyword-first sourcing misses them.

What does "dual-stack contributor" mean and why does it matter?

A dual-stack contributor is an engineer who has merged code to both vLLM and SGLang, the two dominant open-source LLM serving frameworks. It matters because these people have proven they can operate across competing designs (PagedAttention versus RadixAttention, different scheduler models) rather than being married to a single codebase. Research from inclusion-ai.org counts 194 of them globally, which is roughly 30% of SGLang's total contributor base and a much smaller share of vLLM's, and this is the tightest publicly identifiable expert pool in the entire inference-cloud labor market.