Where defense tech gets its hardware engineers.

We didn't filter for hardware titles up front. We classified every engineering hire and let the data draw the line.

Within the confirmed-hardware band: 1,444 hybrid (mechatronics, embedded, integration and test) and 1,307 physical. The unresolved band matched no keyword on either axis and was previously folded into software without being shown; it is broken out here because it is large enough to move the answer.

What it shows: the flow from The Arena's full 468-company tracked universe down to the 2,751-person hardware cohort this issue is built on, plus how that cohort split from the raw classification.

So what: the honest answer is a range, not a point. Hardware is at least 48.2% of engineering hires at these 35 companies and at most 73.7%; software is at least 26.3% and at most 51.8%. The width comes from 1,454 hires (25.5%) whose titles matched no discipline keyword — “Engineer,” “Chief Engineer,” “Launch Engineer,” “Mission Operations Engineer.” A spot check found at least 168 of them carrying an explicit hardware term, so the unresolved band leans hardware rather than sitting neutral. What we can say without a range: at a company whose identity is building physical hardware, a very large share of engineering hiring is still software, and every downstream number in this issue runs on the 2,751-hire confirmed-hardware floor.

Caveats: classification is rule-based and imperfect at the margins. Hybrid (mechatronics, embedded, integration and test) sits between the two worlds by design. Titles that matched no keyword default to digital in the classifier, which makes the software share an upper bound and the hardware share a lower bound — the reverse of what an earlier version of this page said.

Who counts as a hardware engineer. Mechanical, electrical, propulsion, thermal, RF, avionics, naval architecture, manufacturing/production, integration and test, controls.

Window: trailing 12 months, genuine new external hires only, interns and internal moves excluded.

SOURCE / METHOD — 35 companies, 5,709 engineering hires classified on 3 axes (role family, artifact, lifecycle) via Crustdata (cached). Feeder index = a source's share of hardware hires divided by its share of the eligible engineer population.

SpaceX and Anduril alone account for about 60% of the 2,751 hardware hires. The mid-caps beneath them are smaller in volume but carry deeper defense pedigree, as section 10 shows.

What it shows: hardware-engineering hires at every one of the 35 companies in scope, trailing 12 months. Top 10 shown by default; "View all 35" reveals the rest.

So what: SpaceX and Anduril operate at an entirely different order of magnitude than the other 33 combined, hiring in the hundreds while everyone else hires in the dozens. Treating "hardware hiring" as one undifferentiated market is misleading here, size-normalizing (next section) is what actually reveals which of the smaller companies are punching above their weight.

Caveats: LinkedIn profile updates lag real start dates by weeks, so the most recent month or two is undercounted for every company here.

By raw headcount, Northrop Grumman feeds more engineers into this hardware cohort than SpaceX does, 101 hires to 69 — both counted at the parent entity, so Northrop Grumman Corporation and Northrop Grumman Mission Systems roll together, as do SpaceX and Space Exploration Technologies. Adjust for company size and Northrop falls to the middle of the pack at 3.88×. The real disproportionate feeders are commercial-space firms a hundredth its size.

What it shows: all 22 companies we could size (published engineering headcount), ranked by how likely their engineers are to move into defense-tech hardware, adjusted for how many engineers they have. Every row carries its observed hire count, its denominator, and a 95% Poisson interval.

So what: commercial-space and new-space firms sit at 4 to 37 times their size; some primes over-index (GD, Northrop, L3Harris, Boeing) and some don't (RTX, Lockheed); every one of the 6 Big Tech firms we could size sits between 0.1 and 0.4 times, the mirror image of Issue 01, where Amazon was the number one software feeder. If you're building a hardware team, you're competing for talent with SpaceX, Blue Origin, Rocket Lab, Relativity and Tesla, not with AWS, and a blanket "go after the primes" strategy won't get you there.

Caveats: the denominator is all-engineering headcount, including non-hardware engineers, so lifts are directional. A feeder index requires a published headcount, which only these 22 of the many source companies in this data have; the raw-count table below has the fuller feeder list, including sources with no index. Lockheed Martin sits at 0.09× on 3 confirmed direct hires (95% CI 0.02–0.26×), up from a hard 0.0× in the original 26-company cut. Hires whose named prior employer was ULA, the Lockheed/Boeing joint venture, are now genuinely excluded from every company's count rather than assigned to either owner — an earlier version of this page stated that exclusion but the calculation was still assigning all 7 ULA hires to Lockheed, which is why Lockheed previously read 0.29× on 10 hires. Two other name-matching errors are also corrected: 6 hires from firms with “metal” or “Rheinmetall” in the name had been counted as Meta (0.39× → 0.23×), and 2 from “intelligence”-named organizations as Intel (0.36× → 0.32×). Apple shows 0 confirmed direct hires: all 12 rows previously credited to it carry the staffing-vendor label “SBT for Apple,” which is not Apple employment.

Even the number one named feeder is under 4% of all classified hardware hires. In this dataset, hardware talent doesn't concentrate in a handful of famous names the way software's did in Issue 01. It's drawn from a very long tail.

What it shows: the named sources of hardware talent by raw hire count (top 10 shown, "View all 17" for the rest), reconciled against the full 2,751-hire base so every share you see is a true percent of the whole dataset, not of some smaller shown subset.

So what: even Northrop, the single largest named feeder, is under 4% of all hardware hiring, and 40% of every hardware hire in this dataset came from a company that sent exactly one person. Software had an Amazon-sized anchor tenant; hardware doesn't have an equivalent. If you're building a hardware team around 4 or 5 marquee feeder companies, you'll miss most of the available talent.

Caveats: this table is raw volume only, so it will overweight simply-large employers; the 22 companies with a published engineering headcount get a size-adjusted lift instead in section 3. "Martin Marietta Energy Systems," a defunct LinkedIn label for an Oak Ridge/DOE site, and "ULA, United Launch Alliance," the Lockheed/Boeing joint venture, were both removed as data artifacts. Counts are rolled up to the parent entity so that this table and the feeder index in section 3 agree: Northrop Grumman Corporation plus Northrop Grumman Mission Systems is one row at 101, Boeing plus Millennium Space Systems plus Boeing Defense is one row at 71, SpaceX plus Space Exploration Technologies is one row at 69. An earlier version of this page mixed exact profile labels here with rolled-up parents in section 3, which is why the two sections disagreed by a hire or two. 62 hires carry one of the two artifact labels and 11 have no resolvable prior employer at all, so 73 hires are excluded from this table — an earlier version said 66. Of the artifact rows, 55 are the “Martin Marietta Energy Systems” label; those people remain in the 2,751-hire cohort and in every other cut in this issue, they are only withheld from the feeder tables.

Across the companies we tracked, the primes don't own hardware. They split. Some over-index hard, some barely feed at all.

What it shows: the six legacy primes, plotted on either side of the 1.0× line. Right if they over-index as a hardware feeder, left if they under-index.

So what: if you're courting prime-alumni talent, General Dynamics, Northrop, L3Harris and Boeing are where it actually moves; don't assume Lockheed talent is in play just because it's a prime. The same pattern held in Issue 01's software data too, where Lockheed contributed zero of 956 hires, so this isn't specific to hardware roles.

Caveats: per-firm lifts rest on all-engineering denominators. A prime that under-indexes overall may still over-index in one specific discipline, such as naval or RF, that this issue doesn't break out. Lockheed Martin's 3 confirmed hires here, out of the 35-company panel's 2,751 hardware hires, are not enough volume to call this a real trend, but it is no longer a hard zero as it was in the original 26-company cut. Hires whose named prior employer was ULA, the Lockheed/Boeing joint venture, are excluded from every company's count rather than assigned to either owner; in an earlier version of this page they were still being counted for Lockheed, which inflated its index to 0.29×.

Across the 35 companies we tracked, hardware and software hires come from a strikingly similar, commercial-dominated pool. Hardware just leans a little more toward defense and a lot less toward Big Tech.

What it shows: across all 35 companies, where hardware hires worked before compared with where software hires worked before.

So what: combine the defense-adjacent categories and hardware pulls 19% of its hires from primes and defense startups combined, versus 13% for software, about 1.5 times as much, while software pulls nearly twice as much from Big Tech (9% vs 5%). That's a real tilt, but it's a tilt, not a different candidate pool: both lists are roughly 57 to 62% plain commercial. If you're standing up both teams, the same recruiting funnel works for both, just tuned differently at the margins. Inside the stricter 6-company panel that also appears in Issue 01, the pattern holds (see appendix).

Caveats: "prior world" is the immediate previous employer; categorization is rule-based; shares are of the classified cohort.

The hardware cohort is mechanical, systems/integration-and-test, and manufacturing first, the disciplines that turn a design into a flying, floating, shooting product.

What it shows: all 2,751 hardware hires across the full 35-company panel, grouped by engineering discipline.

So what: "Robotics" gets the headlines, but it's the smallest named discipline here, just 37 hires, barely 1% of the cohort, smaller even than Naval/Marine. The actual volume is in unglamorous, build-and-qualify disciplines: Mechanical/Structures alone (733 hires) outnumbers Robotics 20 to 1. If your mental picture of defense-tech hardware hiring is a robotics wave, the data says it's a mechanical-engineering and manufacturing wave instead.

Caveats: discipline is inferred from job title. Systems/Integration & Test is a broad bucket and was spot-checked for accuracy. 143 hires (5.2%) carry a hardware artifact term but no discipline-specific keyword (for example "Chief Hardware Architect," "Hardware Development Engineer") and are shown as their own unresolved bar rather than distributed across the named disciplines. An earlier version of this page showed the nine named disciplines only, with that residual redistributed into them by a pass that was not recorded and could not be reproduced; the bars above are the reproducible stored counts, so several read lower than before (Mechanical/Structures 826 to 733, Propulsion/Thermal 313 to 282, Controls/GNC 156 to 129). The ranking and the conclusion are unchanged.

Only about a quarter of the hardware hires we tracked came straight from defense. Trace the full career and over half have touched it.

What it shows: the share of hires with defense/military experience, measured two ways: last job only versus entire career, hardware versus software.

So what: the boomerang pattern means a search that filters only for "currently at a defense company" will systematically miss the more experienced half of this pool. The strongest hardware candidates often sit at a commercial firm today with a prime or a service tour two jobs back that a title-only search would never surface.

Caveats: "ever touched defense" includes short and early stints; it is not the same as a career defense specialist. Rule-based classifier.

In this dataset, hardware hiring is far more concentrated by state than software's was, and Texas's pace has nearly doubled over the past six months.

What "location" means here: each hire's own current LinkedIn-listed location, where the person says they are today. It is not the hiring company's HQ and not the prior employer's HQ. Coverage: 2,466 of 2,751 hardware hires (90%) list a specific location. The remaining 285 (10%) list none and are excluded from the map and tables below. They are not silently dropped from the report, they simply don't appear in a geography cut.

What it shows: the 8 U.S. states with a meaningful hardware-hire count in this dataset, each bubble sized and labeled with its own count, out of 2,349 hires that resolve to a U.S. state.

So what: California alone accounts for 49% of every hardware hire with a known state, essentially a coin flip that lands on one state, with Texas a distant second at 14%. Hardware hiring is inherently local, and the talent pool for a given discipline outside these 8 states is thin enough that relocation, not remote work, is usually the real lever.

Caveats: self-reported LinkedIn location, not employer HQ. States not shown here account for the remaining U.S. hires; New York edges out Virginia for the 8th spot, on 33 hires to 31.

What it shows: the 117 hardware hires, 4% of everyone with a known location, who list a location outside the United States, grouped by country and by the company driving that cluster.

So what: this is not U.S. companies staffing engineering roles abroad. It's almost entirely explained by four of the 35 tracked companies that are themselves headquartered outside the U.S. (ICEYE in Finland, with satellite offices in Poland and Spain; Quantum Systems in Germany; Harmattan AI in France; DroneShield in Australia), plus Anduril's real UK and Australian offices. Don't read this as an opening to hire hardware engineers internationally for a U.S.-based role: most hardware engineering positions at U.S. defense-tech companies are export-controlled and require U.S. citizenship, so this cluster reflects which companies happen to be foreign-headquartered, not a generally available international talent pool.

Caveats: self-reported LinkedIn location; sub-national regions (for example Uusimaa, New South Wales, Bavaria) are rolled up to their country.

What it shows: the top named metros among the 1,957 hardware hires (71% of the total) who listed a specific city.

So what: these 10 metros account for 761 hires, 39% of everyone with a listed city, but only 28% of the full 2,751; the rest is spread thin across 493 other named places. Long Beach enters the top 10 after the merge, pushing San Diego (32 hires) just outside it.

Caveats: city-level detail is thinner than state-level; treat this as directional context alongside the map above, not a full accounting.

What it shows: hardware-hire volume by state, comparing the back half of 2025 (August–December) to the first half of 2026 (January–June).

So what: Texas's growth is real and broad-based across many employers, roughly doubling in pace; Ohio's and Michigan's are real but driven almost entirely by one company standing up a specific site, not a new regional hub yet. If you're an employer opening a site in a state like Ohio or Michigan, you're not stepping into an existing hardware labor market, you're building one from close to zero, and should expect to relocate talent rather than hire it locally at first. If you're an engineer weighing where the momentum is, Texas is now the deepest bench outside California.

Caveats: July and August 2026 are excluded from both windows. The most recent month or two in any LinkedIn-based hiring dataset are typically undercounted, because new hires haven't yet updated their profile with their new employer, so a naive full-window comparison would make hiring look like it's slowing down when it almost certainly isn't.

The 20 hardware-native mid-caps lean even harder on commercial-space feeders and carry deeper defense pedigree than the 6 giants. This isn't an artifact of scale.

What it shows: the same two cuts, feeder propensity and defense lineage, broken out separately for the 6 giants versus the 20 smaller, hardware-native companies.

So what: the pattern from the rest of this issue isn't an artifact of Anduril and SpaceX's scale, if anything it's stronger at companies closer to the mission. If you're a hardware-native company rather than one of the two giants, commercial-space hires and defense-boomerang candidates are an even better bet for you than they are at SpaceX or Anduril's scale.

Caveats: the mid-cap panel is 20 companies, several small, so per-firm figures are directional. Both cuts use the same classification and window as the rest of this issue.

Everything needed to check this page against its own data.

Aggregate tables and claim ledger: every chart and table on this page, plus the reproducibility ledger and a de-identified row-level extract, are downloadable as spreadsheets at /data/arena-insights-data-pack-v1.zip. Person names and profile URLs are not included and are held privately for audit.

What changed in this revision: the section 1 split is now reported as three bands and a range rather than a single hardware/software percentage; the feeder index carries observed counts, denominators and 95% intervals, and its true denominator is stated; ULA, "metal"-named and "intelligence"-named false matches are removed from the index, moving Lockheed Martin to 0.09×, Meta to 0.23×, Intel to 0.32× and Apple to zero confirmed direct hires; sections 3 and 4 now use the same parent-entity roll-up, so Northrop reads 101 and SpaceX 69 in both; the discipline bars are the reproducible stored counts with the unresolved residual shown; and the inherited 79% software figure from Issue 01 is corrected to 78%. No conclusion in this issue changed direction.

Scope: 35 defense-tech manufacturers. 5,709 engineering hires. Trailing 12 months. 2,751 classified into the hardware cohort (physical/hybrid artifact).

• Company selection: the panel combines a hand-selected 26-company set (a 6-firm cohort also tracked in Issue 01, plus a 20-firm hardware-native coverage panel) with a 9-firm, data-driven expansion: the highest-volume hardware and hybrid hirers by trailing-12-month engineering hire count, drawn from the 238 hardware/hybrid companies in Arena's tracked 468-company universe that weren't already in the hand-picked 26. The second cut was built to test whether the hand-picked panel's conclusions were an artifact of curation; they weren't.

• Which chart uses which slice: sections 2, 4, 6, 7, 8 and 9 (scale, raw feeders, hardware vs. software, disciplines, lineage, geography) run on the full combined 35-company, 2,751-hire hardware cohort. Sections 3 and 5, the feeder index and the primes chart, are ratios of two shares and can only be computed for hires whose prior employer resolves to one of the 22 firms with a published engineering headcount — 588 of the 2,751, stated on the chart itself. Section 10 is a deliberate split of the 6 giants from the 20 mid-caps, to test whether the pattern is scale-dependent. No chart draws on an unstated subset; where a chart uses less than the whole cohort, that count is shown on the chart. An earlier version of this page said sections 2 through 9 all ran on the full 2,751; the feeder index never did.

• Feeder index equals a source's share of hardware hires divided by its share of the eligible engineer population. Denominator is all-Engineering headcount per source, so lifts understate hardware-specificity for software-heavy firms and are directional, not exact. Only sources with a published engineering headcount, 22 of the many feeders in this data, get an index; the raw-count table in section 4 has no such requirement. The 2 hires whose prior employer was NVIDIA are excluded from this index, same as any other source we never pulled a headcount for. Source names are matched to a parent entity with an ordered pattern list; the patterns are anchored to avoid the false matches an earlier version of this page carried ("metal"/"Rheinmetall" matching Meta, "intelligence" matching Intel, and ULA matching Lockheed). Every feeder row in sections 3, 4 and 5 now carries its observed hire count so a reader can see which rankings rest on real volume and which rest on single digits.

• Three-axis classification (role family, digital/physical/hybrid artifact, lifecycle stage) is rule-based, run on each hire's job title. Role family and artifact are assigned together in a fixed order: titles containing mechanical, electrical, RF/antenna/radar, propulsion/thermal, structural, naval/marine, or materials/welding terms are tagged physical; titles containing avionics, GNC/controls, robotics/mechatronics, embedded/firmware, manufacturing/production, or integration-and-test/systems-engineer terms are tagged hybrid; titles containing machine-learning/data, platform/infrastructure, security, forward-deployed, or software/developer terms are tagged digital. A title matching none of these defaults to digital under an "Other/Unresolved" role family. That default runs one way, so the digital count in section 1 is a ceiling on software's true share, not a floor — and the hardware count is a floor, not a ceiling. It is a large default: 1,454 of the 2,958 hires counted as digital (49%) matched no software keyword and arrived there only by falling through, and a hand check found at least 168 of them carrying an explicit hardware term (Structural Engineer, Avionics Engineer, Tooling Engineer, Harness Design Engineer). Section 1 therefore reports three bands and a range instead of a single split. An earlier version of this page described the same rule as producing a floor on software; that was backwards. Lifecycle stage is a separate pass over the same title, looking for research, integration, test, production, or deployment/operations language, and otherwise defaulting to design/build. Low-confidence and high-impact buckets, including the "Other/Unresolved" catch-all and the broad "Systems/Integration & Test" bucket, were spot-checked by hand. A known LinkedIn data artifact, "Martin Marietta Energy Systems," an old label for a DOE/Oak Ridge site, was removed from the feeder tables, as was "ULA, United Launch Alliance," Lockheed Martin and Boeing's joint venture, which we excluded rather than assign to either owner.

• Career-history claims use the full work history, not the immediate prior employer alone. "Touched defense" means a prior employer at any point in that person's career fell into one of four categories: legacy prime (for example Lockheed Martin, Northrop Grumman, RTX/Raytheon, L3Harris, General Dynamics, BAE Systems, General Atomics, Leidos, Booz Allen Hamilton, SAIC, Sikorsky, plus shipyards such as Huntington Ingalls and General Dynamics Electric Boat, and OEMs/tier-1 suppliers such as GE Aerospace and Honeywell); neo-prime or other defense startup (for example Anduril, Palantir, Shield AI, Saronic, Skydio, Scale AI, Applied Intuition, and every one of the 35 companies in this issue's own tracked universe, plus any employer whose LinkedIn industry tag reads "Defense and Space Manufacturing" even without a name match); military and government (the named service branches, the Department of Defense, DARPA, NASA, and national labs and federally funded R&D centers such as Sandia, Los Alamos, Lawrence Livermore, Johns Hopkins APL, and JPL). A prior employer that did not fall into one of those four categories does not count as "touched defense" under this rule, even if that company does some defense-adjacent or government-contract work; commercial firms, Big Tech (Amazon, Microsoft, Meta, Google, Apple, NVIDIA, Tesla), and academic institutions are excluded from the defense count by design, so this is a conservative measure that will undercount defense exposure at the margins.

• LinkedIn lags real start dates by weeks, so the most recent months are undercounted. Signal reflects current employees only, so departed hires are invisible.

• Source: a combination of web search, AI, Crustdata, and manual validation were used to develop this article. Person-level rows are held privately for audit and are not published.

Every one of the 35 companies in scope. Full auditable list; nothing in this issue draws from a company not shown here.