Together AI's 500 MW Bet: Source OCI Refugees, Not ML PhDs

Together AI closed an $800M Series C at $8.3B on July 1, 2026, with 500+ megawatts of compute committed separately and a projected 50x infrastructure expansion over five years. Every AI recruiter in the Bay Area is now hunting the same ML PhDs. They are hunting the wrong pool. The binding constraint here is megawatts and the people who wire them, and Oracle handed that exact profile to the open market 90 days earlier.

What Together AI actually needs to hire

Together AI needs data center, critical facilities, power, and inference-serving infra engineers, not more ML researchers. The company already ships FlashAttention-4 and Together Megakernel; its research bench is not the bottleneck. The bottleneck is turning capital into cooled, powered, GPU-populated floor space.

Do the math on the Series C:

• 500+ MW committed compute, capitalized separately from the $800M
• ~300 current employees, 55 open roles as of March 2026
• ~1.67 MW per current employee, vs. hyperscaler norms of 0.5 to 1.0 MW per employee
• 50x infrastructure footprint growth targeted over five years
• $1.15B annual bookings last quarter, valuation up from $3.3B just 16 months prior

That MW-per-head ratio is the tell. Together AI is roughly 2 to 3x underweight on physical-infra headcount before the new capacity even lands. Sourcers reading the press release as "AI hiring" will index on Kaggle grandmasters. The actual req sheet skews toward staff-level SREs, RDMA network engineers, GPU provisioning leads, and site power specialists.

1.67 MW

Why the Oracle cuts are the arbitrage

Oracle laid off approximately 30,000 employees on March 31, 2026, roughly 18% of its global workforce, and OCI engineering took a disproportionate share. That is the exact pool Together AI needs, and most recruiters are misreading it as commodity support headcount. It is not.

Three facts change how you should source it:

1. Oracle is spending $50B on cloud infrastructure this year while firing the cloud engineers who build and maintain it. The paradox is the arbitrage.
2. A Computer Weekly-cited Oracle security manager described the algorithm as targeting "high-level individual contributors and mid-level managers, especially those with outstanding stock options." This is a senior pool, not a junior one.
3. OCI support and customer success absorbed an estimated 6,000 to 8,000 role cuts, but the valuable subset for Together AI is the ~15 to 20% who touched bare-metal GPU provisioning, RDMA networking, Slurm, and multi-tenant Kubernetes, not ERP-adjacent cloud roles.

The window is short. Tech-Insider reported that AWS, Microsoft, and Snowflake began aggressive outreach to OCI specialty talent "within hours" of the layoff email. Fireworks AI, Baseten, RunPod, and CoreWeave are on the same list. If your outbound loop is measured in days, you are behind.

The pools, numbered

The addressable supply looks like this, pulled from Refolk's index of professional profiles. Only one of these columns matters, and it is not the one recruiters usually stare at.

Two numbers to sit with. First, 370. That is your entire realistic Together AI shortlist for OCI-flavored senior cloud engineers, and it is shared with every hyperscaler and inference startup in the market. Second, 2,839. That is your data center engineering pool, and it is where the actual megawatt problem gets solved. The ratio of 7.7 is why the sourcing strategy has to be two-track, not one.

370

Filter on RDMA and Slurm, not on "OCI"

The word "OCI" on a LinkedIn profile is a filter, not a moat. It surfaces 392 US profiles but most of them are ERP-adjacent, Fusion Cloud, or Oracle Database Autonomous work that does not transfer to a GPU cluster. Filter on the substrate instead.

The keyword stack that actually predicts fit for a Together AI infra role:

• Bare-metal GPU provisioning
• RDMA over Converged Ethernet (RoCE) or InfiniBand fabric engineering
• Slurm, Kubernetes with device plugins
• Multi-tenant Kubernetes and topology-aware scheduling
• Liquid cooling and medium-voltage power distribution
• Data center network fabric at the pod level

This is the gap Refolk closes for infra sourcing: instead of Boolean-stacking "OCI AND RDMA AND (Slurm OR Kubernetes) NOT Fusion" and getting stale profiles, you describe the person in plain English (a staff cloud engineer who ran multi-tenant GPU clusters at OCI, comfortable with RoCE and Slurm, not ERP-adjacent) and get a ranked shortlist that survives contact with a hiring manager.

The second pool nobody is sourcing: Schneider Electric alumni

Together AI's cap table quietly tells you where the second pool sits. Schneider Electric's SE Ventures joined the Series C alongside NVIDIA, Aramco Ventures, Vista Equity Partners, General Catalyst, and Emergence Capital. Pegatron is also on it. That is not just capital. That is a hiring pipeline signal for power and hardware integration, and almost no AI recruiter is reading it.

The mining targets that fall out of the cap table:

• Schneider Electric power engineers, especially those who have done hyperscale deployments
• QTS Data Centers critical facilities engineers, which Refolk's index surfaces as a top employer for the 2,839-person DC pool
• Pegatron hardware integration engineers
• NVIDIA deployment engineers, the ones who install rather than design

The cap table is a hiring plan. Schneider Electric did not write a check to help you find ML PhDs.

Together AI does not have to poach these people cold. Investor-adjacent outreach is warmer than a recruiter InMail, and the story ("we are the company your fund just backed to deploy 500 MW") writes itself. Sourcers should treat every logo on the July 1 announcement as a first-degree pool to enumerate.

Geography: Austin beats the Bay

Austin, not San Francisco, is the highest-yield zip code for this specific search. Refolk's index shows Austin as the top concentration for both OCI talent and platform engineers in the sampled cohort. The reason is structural, not vibes.

• Oracle's HQ is in Austin, which means the March 31 layoffs concentrated there
• Data Center Dynamics reported OCI team cuts hitting both US and India, and the US share sits heavily in Texas
• Many affected engineers have Texas roots and no appetite to relocate
• Together AI already has cloud footprint decisions to make in Texas power markets

If you are running an outbound program for Together AI, Austin should be a heavy majority of your sourcing volume, and the Bay Area should be treated as the second market, not the first. Every other AI-infra recruiter is inverting that ratio and paying for the mistake.

Speed of loop is the whole strategy

The 370-person addressable OCI pool is being drained by hyperscalers running sub-72-hour loops. If your process is a two-week phone screen funnel with a panel loop in week 3, you will lose. This is where Together AI's recruiting posture has to change materially from a normal Series C hire.

The concrete moves that actually compound:

1. Same-day reach on any surfaced profile. Not next business day. Same day.
2. Skip the recruiter screen. Send the hiring manager to the first call, ideally within 48 hours.
3. Compress the loop. Onsite in a week, offer in two.
4. Pre-clear comp bands. Do not open a req you cannot fund at staff-level infra rates.
5. Named-list sourcing over inbound. Inbound is where you get the ex-Fusion engineer. Named-list is where you get the RDMA lead.

The named-list piece is where sourcing tools earn their price. Enumerating "the ex-OCI engineers with GPU cluster experience, ranked" is exactly the kind of query that used to take a sourcer days of Boolean and LinkedIn Recruiter seat-swapping. Refolk collapses that to a paragraph of English and a shortlist you can outreach the same afternoon, which matters when the alternative is losing candidates to a Microsoft recruiter who called at 9 a.m.

What to ignore

Together AI's public-facing hiring narrative will emphasize research prestige, FlashAttention-4, Megakernel, and the NVIDIA partnership. That is recruiting marketing surface. The actual staffing plan is the people who know how to keep a GPU cluster on power and on network.

Signals to deprioritize when sourcing for this account:

• Kaggle rank and ML PhDs without systems experience
• Generic "cloud engineer" titles without a GPU or HPC substrate
• Oracle profiles heavy on Fusion, ODA, or Autonomous DB

Signals to overweight:

• Recent departure from Oracle OCI in 2026
• AWS, Microsoft, Meta, or NVIDIA alumni from the infra org, not the product org
• CoreWeave and QTS experience
• Schneider Electric alumni with hyperscale deployment history

The Together AI hire is a data center hire wearing an AI logo. Source it that way and the 500 MW math starts to work. Source it as an AI hire and you will spend six months explaining to the CEO why the racks are late.

FAQ

How many ex-Oracle OCI engineers are actually available for Together AI?

Refolk's index shows 392 US-based profiles with OCI in the headline or experience, of which only 22 are still at Oracle, leaving roughly 370 already out or in transition. The valuable subset for Together AI is the estimated 15 to 20% of that pool with bare-metal GPU, RDMA, or Slurm experience, so realistically 55 to 75 engineers are the direct fit. That is a small enough pool that a named-list sourcing approach beats any inbound funnel.

Why is Together AI's Series C a data center story rather than an AI story?

Because the compute commitment is 500+ MW, capitalized separately from the $800M raise, and the current 300-person headcount cannot physically deploy that footprint. The ratio of 1.67 MW committed per current employee is 2 to 3x the hyperscaler norm, meaning the binding constraint is critical facilities, power, and cluster SRE headcount, not model research. FlashAttention-4 and Together Megakernel already exist; the racks do not.

Which competitors are drawing from the same OCI pool?

AWS, Microsoft, and Snowflake began outreach to OCI specialty talent within hours of the March 31 layoff notification, per Tech-Insider reporting. Inference peers Fireworks AI, Baseten, RunPod, and CoreWeave are hunting the same profiles. Sub-72-hour outreach loops are the only viable posture; a two-week screening funnel will lose every candidate to a faster competitor.

Should sourcing focus on the Bay Area or Austin?

Austin, decisively. Oracle's HQ is in Austin and the layoffs concentrated there, most affected engineers have local roots and low relocation appetite, and Refolk's index confirms Austin as the top US region for both OCI talent and platform engineers in the sampled cohort. A healthy Together AI outbound program should treat Austin as the primary market, with the Bay Area as the second.