SambaNova Just Doubled to $11B. The 77 It Cut in April 2025 Are the Prize.
SambaNova raised $1B on July 8 at $11B. The real sourcing list is the ~77 alumni cut in the 2025 training pivot, not the 397 still there.
SambaNova closed the first tranche of a $1B Series F at an $11B post-money on July 8, 2026, led by General Atlantic. Every chip reporter is writing about SN50 throughput and the Intel entanglement. If you recruit AI silicon talent, that's the wrong story. The interesting list is not the ~397 people still at SambaNova. It's the alumni graph, and specifically the 77 engineers WARN'd out in April 2025 when the company pivoted away from training.
You have roughly six months before Cerebras, Groq, and Nvidia's inference org finish mapping the same graph.
Why the roster is the wrong list
A $1B round at an $11B valuation, five months after a $350M Series E, with IPO-in-2027 chatter attached, is a retention event. Not a hiring event. Not for the people already inside.
Everyone on the current roster is about to get a refresh grant. The founders (Rodrigo Liang, Kunle Olukotun, Chris Ré) are going to spend Q3 stapling equity to the top 50 engineers because their competitors, especially Cerebras (public, ~$49B cap) and Groq ($6.9B, $750M in Sept 2025), can wave real liquidity. Anyone poachable from the current 397 was poachable in May, before the term sheet leaked. That window is closed until roughly month twelve, when the first refresh cliff hits.
The alumni cohort has none of that friction. They already left. They're vested or written off. They know the RDU stack cold. And most of them are exactly one recruiter conversation away from moving again.
The April 2025 cohort is a competitor gift
Read SambaNova's own April 2025 statement carefully. The company said the cut aligned to "the transition we've seen from model training to fine-tuning and inference." Translated: they fired their training people.
Training-scale silicon and systems engineers are the single hottest profile in the market right now. Cerebras still trains. AWS Annapurna/Trainium still trains. Nvidia's DGX Cloud group still trains. Meta MTIA and Google TPU teams still train. SambaNova's "pivot" handed all of them a warm, pre-cleared, chip-native training cohort with WARN paperwork stamped between April and July 2025.
If your sourcing filter is "SambaNova, current employee," you are looking at the wrong two hundred people. Filter for "SambaNova, tenure ended April 2025 through July 2025." That's the layup.
The Boolean also has to encode SambaNova's proprietary vocabulary or LinkedIn's fuzzy title matching will bury the good profiles under generic "ML Engineer" noise. Hard-code these terms into every search: RDU, Reconfigurable Dataflow Unit, SambaFlow, SN10, SN40L, SN50, SambaRack, SambaCloud, composition of experts. Profiles with two or more of those tokens are almost always real. Profiles with none are almost always mislabeled.
This is the exact kind of query where plain-English sourcing beats LinkedIn's own filters. You can describe the human ("ex-SambaNova compiler engineer who left in the 2025 layoff, now at a stealth startup or Nvidia") to Refolk and get back a ranked shortlist across GitHub, LinkedIn, and the open web without hand-tuning Boolean strings around six proprietary acronyms.
Compiler and dataflow first. RTL second.
Not all SambaNova alumni are equal. The most defensible skill inside that company was never chip design. It was the compiler and dataflow layer that maps ML workloads onto reconfigurable silicon.
That skill is portable in a way RTL and verification are not. A SambaFlow kernel author can walk into Groq's TSP compiler team, Tenstorrent's Buda stack, MatX, Etched, Lightmatter, or d-Matrix and be productive in a sprint. A pure RTL engineer with SambaNova on the résumé is competing against a much larger pool from Nvidia, AMD, Apple, and every hyperscaler silicon group.
If you're building a shortlist, weight it like this:
- Tier 1: Compiler, kernel, runtime, and "composition of experts" engineers. Bonus for MLIR, IREE, or TVM commit history on GitHub.
- Tier 2: ML hardware architects and performance engineers who published on RDU throughput or SN40L numbers.
- Tier 3: Field/forward-deployed engineers who ran the Argonne, Lawrence Livermore, SoftBank, or TEPCO deployments. These are your solutions-architect and named-account hires, not your core research pipeline.
- Tier 4: Pure RTL, DV, physical design. Still hireable, but not the differentiator.
The pivot to inference turned a training-org purge into a training-talent subsidy for every competitor with a training roadmap.
The alumni graph you should actually build
Start with the founder ring. Kunle Olukotun's Stanford Pervasive Parallelism Lab and Chris Ré's DAWN Lab (now Hazy Research) are the two academic feeders that produced most of SambaNova's early technical hires. Anyone with a PPL or DAWN affiliation who ever interned or joined full-time is either at SambaNova, at a SambaNova-adjacent competitor, or at a lab that will produce the next one. That is your evergreen watchlist.
Then add the Sun/Oracle SPARC and Niagara diaspora. Olukotun came out of Afara, Liang led processor work at Sun and Oracle. A meaningful slice of SambaNova's senior silicon bench came with them. When those people leave SambaNova, they don't post about it. They just show up at a new employer six months later. LinkedIn's "started a new position" signal is late by design.
Then layer the deployment customers. SoftBank got the first SN50 deployment. JPMorgan Chase is the newest named on-prem inference account. Argonne, Lawrence Livermore, TEPCO Systems, Hugging Face, and Meta all show up in SambaNova's customer graph. Alumni frequently land at their former customers as field engineers or solutions architects. If you're hiring a forward-deployed engineer for your own inference product, those companies are the second-order alumni pool.
Finally, watch Intel. Intel tried to buy SambaNova at ~$1.6B including debt in December 2025. Talks stalled, converted to a partnership plus a ~9% stake, and now the valuation is roughly 7x that offer. That entanglement means SN50 systems co-develop with Xeon, which means engineer movement between the two organizations is going to be bidirectional and unusual. Ex-Habana and ex-Gaudi Intel engineers should be on the same shortlist as SambaNova alumni. They're adjacent supply, and most sourcers won't think to include them.
Read Glassdoor as a supply forecast
SambaNova's Glassdoor page (48 reviews indexed) has a 36% recommend rate. Reviewers describe reorgs and layoffs "at least once a year." One 2024 reviewer noted the company "brought in virtually no revenue, despite relentless pressure on engineering teams."
Every recruiter I know reads Glassdoor as a culture check. That's fine, but the more useful read here is supply forecasting. A company that reorgs annually produces a fresh alumni cohort every twelve months. This is not a company you campaign against once. It's a company you build a permanent watchlist for. Set the alert. Refresh the list quarterly. Every reorg is a new draft.
The $1B round will not change this rhythm. If anything, it accelerates it, because a Series F at $11B on top of a valuation that was $1.6B in December 2025 creates enormous pressure to grow into the number. Growth-into-valuation pressure at chip companies has always produced org churn. See: pretty much every AI silicon startup with more than $500M raised.
The six-month window, concretely
Here's what a real AI chip engineer recruiting plan looks like in the next 180 days.
Weeks 1 to 2. Pull the April to July 2025 SambaNova departures. Cross-reference against GitHub commits to MLIR, IREE, TVM, and any public SambaFlow-adjacent repos. Expect roughly 40 to 60 strong profiles across compilers and ML performance. Refolk's own index surfaces about 42 profiles at the compilers + CUDA + ML cross-section tied to SambaNova, with a top-of-list skew toward Principal ML Engineer, Chief Architect, ML Hardware Architect, and VP Engineering titles. Alumni cluster in Austin, Santa Clara, and the broader SF Bay Area.
Weeks 3 to 6. Reach out. The message that works is not "we're hiring." It's specific: reference RDU, reference the composition-of-experts runtime work, reference the specific training workload they published on. Generic outreach to this cohort has a sub-1% reply rate. Specific outreach lands in the 15 to 25% range in my experience with adjacent silicon roles.
Weeks 7 to 12. Build the second ring. Ex-Habana Intel engineers, ex-Groq compiler team, Stanford PPL and DAWN alumni currently at hyperscalers. This is where a Cerebras Groq talent map stops being a spreadsheet and starts being a live graph.
Weeks 13 to 26. Wait for the reorg. There will be one. There has been one every year. When it hits, you already have the alumni already contacted, already in your CRM, already warmed up. Your competitors will be running cold Boolean strings against "SambaNova" the day the news breaks.
If you're a founder building an inference chip company right now, the Series F sourcing playbook is not to chase the current SambaNova roster. It's to own the alumni graph before Cerebras, Groq, and Nvidia's inference recruiters finish theirs. Tools like Refolk exist precisely so a two-person recruiting team can run this kind of natural-language search across GitHub, LinkedIn, and the open web without spending three weeks tuning filters against "RDU" and "SambaFlow."
The $11B number is the headline. The 77 is the list.
FAQ
Why not just recruit from SambaNova's current 397 employees?
Because a $1B round at an $11B valuation, five months after a $350M Series E, is going to fund an equity refresh across the entire technical staff. The internal comp gap that made SambaNova engineers poachable last quarter is about to close for six to nine months. You'll get better hit rates on people who already left, especially the April to July 2025 cohort, than on people who just watched their strike price get interesting again.
How do I know if a candidate actually worked on the important parts of the SambaNova stack?
Look for proprietary vocabulary in their profile, GitHub, or talks: RDU, Reconfigurable Dataflow Unit, SambaFlow, SN10, SN40L, SN50, SambaRack, SambaCloud, composition of experts. Two or more of those terms usually indicates real work on the differentiated layer. Zero of those terms, and a generic "ML Engineer at SambaNova" title, usually indicates a peripheral role. Cross-reference with MLIR, IREE, or TVM contribution history on GitHub for the strongest signal.
Which competitors will move fastest on this cohort?
Cerebras and Groq have the most direct architectural overlap and the most obvious training-story fit for the April 2025 layoff cohort. Nvidia's inference and DGX Cloud groups have the deepest pockets and will move on senior architects. AWS Annapurna/Trainium, Google TPU, Meta MTIA, and Microsoft Maia are all live buyers. Tenstorrent, Lightmatter, Etched, d-Matrix, MatX, and Rain AI will target the compiler layer specifically. If you're not one of those companies, your window is shorter than theirs.
What's the single highest-leverage first move this week?
Pull the WARN-tied departures from SambaNova between April 22 and July 2025, filter for compiler, kernel, and runtime titles, and send fifteen specific messages referencing the exact workload each person publicly worked on. That single action, done in a week, will outperform any generic AI chip engineer recruiting campaign you can run in a quarter.