Trace.Space Bets Trinity Can Scale Physical AI Fleets

Trace.Space launched Trinity, a hardware engineering graph that AI agents and humans share. The Riga-founded company wants robotics, aerospace, auto and defence teams to see what a change touches before the next build ships.

CEO and co-founder Janis Vavere is betting the limit on those machines is no longer money. It is whether a team can still name the authoritative design after the prototype has already flown.

Trinity Ties a Change Across Three Generations

Trinity connects requirements, tests, design parameters and product variants in one system, then lets agents read that graph. Rather than copying a whole vehicle, robot or satellite every time a new configuration appears, teams mark what is shared and where each version diverges. When something moves, the product is supposed to show what else is hit.

The company will put that claim on a clock. On October 8, 2026, at 1 p.m. Eastern, Vavere, CTO and co-founder Karlis Broders, and forward deployed systems engineer Matt Maclaine, formerly of Anduril, will change a design parameter on a live product family and watch the impact light up across three hardware generations, with Space Agent finding gaps and an engineer approving the work in minutes.

WHAT TRINITY CONNECTS

• Requirements: The statements of what a product must do and how it must perform, which Vavere calls the core of the IP.
• Tests: Cases, steps, runs and results that have to land back on the item they were meant to prove.
• Design parameters: The numbers and constraints that actually change when a generation or a customer variant is born.
• Product variants: The related but non-identical machines that inherit a shared subsystem and still need a decision on every change.

The architecture is API-first, so teams can hook it to the development and test tools they already run. The company also sells a configurable graph that can take different types of engineering data, with native support for parameters and variants, which is the structure it says older requirements tools never had.

Trace.Space now estimates Trinity covers about 40 percent of the product development lifecycle, up from roughly 10 percent when the company started, and it wants 80 percent. Over the next year it plans to push into modelling, simulation and manufacturing, including bills of materials, ERP and supply chains. Vavere’s longer picture is agents walking those graphs across companies and suppliers, instead of weeks of PDF handoffs.

$18.8 Billion Could Not Buy a Fleet

Crunchbase counted $18.8 billion into robotics startups in 2026 through late June, already more than the $15 billion the same tracker logged for all of 2025. Capital has been easy to find for drones, humanoids and autonomy software. Working fleets have not.

THE MONEY AROUND THE MACHINES

• $18.8 billion: Crunchbase tally of global robotics startup funding in 2026 through late June.
• $15 billion: Crunchbase full-year 2025 total for the same category.
• $4 million: Cherry Ventures-led seed for Trace.Space in February 2025.
• 40 percent: Share of the product lifecycle Trace.Space says Trinity covers now.

Vavere’s line is blunt, and it is the wager underneath Trinity.

The constraint is no longer whether ambitious teams can get funded. It’s whether they can engineer, validate, and manufacture products fast enough to win.

Janis Vavere, CEO and co-founder, Trace.Space

He argues the tooling required for that speed used to exist only inside a handful of advanced hardware companies, staffed by internal platform teams, and that Trinity is how everyone else buys it. Cherry Ventures, making its first Baltics investment, led the $4 million seed round in February 2025 with Outlast Fund and existing backers Nebular, Fiedler Capital and Change Ventures. Dinika Mahtani and Dimitri Sedashev, writing for Cherry, even sketched the company as a possible SAP for the next generation of manufacturing. That is an investor’s ceiling, not a booked result.

Software-defined machines keep changing after they ship. A robot, drone, satellite or vehicle takes a software update that has to be traced through requirements, tests and hardware configurations before the next version goes out, and as fleets grow that tracing is what slows the team. Trinity’s pitch is that agents work inside the same graph, under rules the team writes, and that engineers still approve every change.

The Pentagon Fielded Hundreds, Not Thousands

The cleanest public test of that thesis is not a startup demo. It is Replicator.

REPLICATOR’S TWO-YEAR CLOCK

1. August 28, 2023: The Pentagon unveils Replicator, aiming to field thousands of uncrewed, attritable systems by August 2025, with about $1 billion sought across two fiscal years.
2. August 2025: The deadline arrives. A Congressional Research Service brief records a former defence official putting the number fielded at hundreds.
3. January 21, 2026: CRS updates the same brief, still treating the gap between thousands and hundreds as an oversight problem for Congress.
4. July 2, 2026: GAO’s 24th weapon systems assessment, GAO-26-108457, finds DOD still plans to invest over $2.4 trillion in its costliest programs while the average delivery time past 12 years.

The stated reasons for the miss were familiar. Some systems glitched when they met existing command structures, some were unreliable, some were too expensive or too slow to make in the quantity the program was built around, and the Pentagon struggled to buy software that could command large numbers of different drones at once.

Vavere, in a September 3, 2026 essay, keeps those reasons on the table and then points at the word that makes them expensive. Managing many identical things is logistics. Managing many related but non-identical things is an information problem, and that is the problem he says nobody had packaged in a form Replicator could buy. Field data comes back, a supplier ends a part, a foreign customer needs a different sensor for a different rule, a software update changes a thermal margin and a test nobody re-ran is now wrong, and each of those hits has to be judged against every live variant.

Most teams copy. A new flight test gets its own spreadsheet, a new customer folder gets its own pile, and a base requirement has to be applied fifteen times by whoever remembers. The authoritative file is the one saved last. Between companies it is worse, because the medium is still a document. A battery supplier’s senior engineers spend days reading a thousand-page spec against the previous thousand-page spec to find a ten-centimetre shift that would otherwise force a new robot onto the line. A warship builder described a six-month design review because a two-line change sent seventeen people back through the whole text.

GAO’s 2026 assessment is the aggregate version of that mess. Across 72 programs with usable cost data, 46 reported increases totalling $122 billion. These are organisations that can buy factories. What they could not retrieve, on a given Tuesday, was which version of the design was live. Vavere’s closing line in that essay is the bet in one sentence: the hundred thousand units are missing because the information layer was built for products that never stop shipping.

What a Late Requirements Error Costs

A NASA Johnson Space Center study led by Jonette Stecklein measured how fast that delay turns into money. Set the cost of fixing a requirements error during the requirements phase at 1. The same error costs 3 to 8 in design, 7 to 16 in manufacturing, 21 to 78 at integration and test, and 29 to more than 1,500 once the system is in operation. The paper assumed a hardware and software program with the shape of a large spacecraft, a military aircraft or a small communications satellite.

NASA COST TO FIX A REQUIREMENTS ERROR

Scaling a program, in that arithmetic, is the act of pushing every unreconciled decision into a later phase. If you cannot tell which requirements a variant satisfies, you find out in integration. If a supplier change quietly kills a test, you find out in the field. Faster prototyping makes the curve steeper, because AI now lets a small team generate code, spin CAD and simulation, and dump more versions into the world with the same headcount. Vavere says every release then compounds risk, because systems engineers cannot keep up with the data, and teams stop knowing whether they are still building the right product.

That is the unflattering reading of the physical-AI boom. The same models that shrink a prototype cycle flood the analysis loop. When tests run weekly and the write-up still runs monthly, unread evidence sits in the pile and the picture of the product is always stale.

From Jama Sales Calls to an Agent Graph

Trace.Space was founded in 2022 in Riga by Vavere, Broders and Mikus Krams, with offices there and in the United States. The three did not come from a blank page. Vavere had led sales at Jama Software, sitting with teams that had bought a requirements system and still could not see how a change in one part hit the rest of the product. Broders had implemented Jama and Polarion on large automotive and government programs and hit the same wall from the inside. Krams had scaled technical teams at Lokalise and Chili Piper.

They reached a shared conclusion: another feature on the old tools would not do it. The way engineering information was structured had to change. IBM DOORS, Jama Connect, Siemens Polarion and Teamcenter still dominate regulated hardware, and they still ask for specialist administrators, long implementations and a thicket of connectors to PLM, test and CAD. Those products were built to hold documents and hierarchies. They were not built as a graph an agent can walk.

Vavere has said customers who previously worked at SpaceX told him Trinity resembles the way they handled engineering data there. That is his anecdote, and it is doing a lot of work. If it is true, a seed-stage company is trying to productise an internal advantage that SpaceX-class firms staffed with their best engineers. If it is sales colour, Trinity is another cloud requirements tool with a chat window. The October 8 demo is where that difference has to show up on someone else’s product family, not a slide.

The company lists Lucid Motors, Serve Robotics, Xiphos, TMAP Mobility and StandardX among the engineering teams it works with. It deploys in the cloud, in a virtual private cloud, on-premise and fully air-gapped, and it says the product is certified to SOC 2 Type II and ISO 27001. Space Agent, the company says, can run quality and compliance checks against DO-178C, ISO 26262, IEC 62304 and ASPICE, with every recommendation routed through engineering review before it lands.

An Agent That Lives in the Trace Chain

Space Agent went live in the product on April 2, 2026, after what Vavere describes as three years of versions that were not good enough for real engineering work. A general model pointed at a requirements database gave general answers. An output that is 80 percent right is, in this job, wrong. The shift was to stop bolting a chatbot onto the side of the tool and to put the agent inside the traceability chain, the coverage model and the requirement structure.

Maclaine had been the in-house sceptic. Then, on a recorded walkthrough, he asked for trace suggestions, a feature that had been added days earlier, and the agent came back with 20, each with an accept button in the interface.

I can do what I got my degree for. Actually make things, instead of writing about making things.

Matthew Maclaine, Forward Deployed Systems Engineer, Trace.Space

The company claims core systems engineering tasks run 10 to 100 times faster than by hand: coverage maps in seconds, blast-radius impact when a requirement changes, drafts of new items from parent specs, well-formedness checks against INCOSE and ISO/IEC language rules, and continuous hunting for broken traces. Those numbers are Trace.Space’s, measured on its own workflows. Customers, Vavere said, are already using the product as a harness, calling it through the API and running their own agents against the same graph to do analysis.

That is the part of the bet that is easy to oversell. An agent that suggests traces a human still has to accept is useful. An agent that is allowed to change a safety requirement on a vehicle or a munition is a different product, and Trinity’s own materials keep a human on the approve step. The graph is what makes the agent useful. Without it, the model is searching prose again.

October 8 Puts Trinity on a Live Product Family

The company posted the Trinity pitch on September 29, 2026: record all engineering data in a single graph, relate every requirement, test, parameter and variant, then reason about change impact across the whole thing.

https://t.co/7awmgrhDSQ launches Trinity: the AI harness that makes scaling physical AI possible.

Hardware teams have the ideas, the funding and working prototypes. What slows them down is finding out, weeks too late, what a change broke.

See Trinity live on October 8. pic.twitter.com/Sq7IaM6ShC

— Trace.Space (@tracedotspace) September 29, 2026

Hardware teams, that post said, already have the ideas, the funding and working prototypes. What slows them down is finding out, weeks too late, what a change broke. The live launch keynote on October 8 is the first time outsiders will watch that sequence on a real product family, with Space Agent in the loop and an engineer still on the accept button.

The bet can fail in ordinary ways. Replicator’s miss had several causes, and a prettier graph will not stamp a hull or yield a command system that can fly mixed drones. Incumbent PLM and ALM suites still sit inside the primes, and they will add their own agents. A $4 million seed does not buy the years of trust a flight-safety program requires before it rips out DOORS. Vavere is still asking those programs to treat engineering information as objects with dependents, variation as a property of the data, and analysis as something that runs at the rate the tests arrive.

On October 8, 2026, at 1 p.m. Eastern, the founders will move one parameter on a live family of hardware. The graph will show the blast radius across three generations, or it will not.