Helm.ai Signs US$70 Million in Contracts, Expands Beyond Autos

Helm.ai has signed US$70 million in commercial contracts over the past twelve months for its foundation-model platform, the company said this week, and the more telling detail in the announcement is not the dollar figure but the customer mix behind it: automotive OEMs and Tier 1 suppliers remain the core, but industrial automation and robotics customers, including perception systems for heavy equipment used in open-pit mining, now sit alongside them. For a company that built its reputation almost entirely on automotive perception software, that diversification is the signal buyers outside the car industry should be watching, because it marks the first concrete evidence that Helm.ai's underlying technology transfers to a market built on entirely different operating conditions, dust, vibration, remote sites, rather than city streets.

Helm.ai, led by chief executive Vladislav Voroninski, built its business on what it calls Deep Teaching, a training methodology that learns the structure of the physical world from unlabeled data before a separate stage teaches the system how to act on what it has learned. The company's pitch to buyers has always rested on a specific claim: because the perception layer is trained without hand-labeled data describing every object and scenario in advance, it should need less labeled data overall and generalize to situations it has not explicitly seen, a property that matters more in mining and heavy industrial automation than almost anywhere else, since dust-obscured cameras, irregular terrain, and equipment operating in locations far from any labeling workforce make exhaustive labeled datasets both impractical and expensive to collect.

Why Mining Perception Is a Harder Test Than a City Street

Open-pit mining is, on paper, a more forgiving environment for autonomous perception than a crowded city street: no pedestrians darting between parked cars, no cyclists, no traffic signals to misread. In practice, it is a harsher test of a perception system's robustness for a different reason entirely. Dust clouds thrown up by haul trucks can obscure cameras and lidar for seconds at a time, equipment operates on uneven, constantly reshaped terrain that has no fixed map the way city streets do, and the economic cost of a perception failure is measured in tens of millions of dollars of haul-truck downtime or, worse, a safety incident involving equipment that can weigh several hundred tons fully loaded. A perception stack that merely works well in clear conditions on a paved test track has not been tested against the conditions that actually determine whether mining operators will trust it with an unsupervised haul route.

That is precisely the kind of environment where a training approach built around reducing dependence on exhaustively labeled data should, in theory, have an advantage, since no mining operator has the kind of labeled-data infrastructure that automotive OEMs have spent a decade building for autonomous driving programs. Whether Helm.ai's Deep Teaching methodology actually delivers that advantage in the field, rather than simply in the company's own framing of its technology, is not something this announcement settles on its own. Helm.ai named no specific mining customer, no site, and no performance benchmark comparing its system against the lidar-and-labeled-data approach that incumbent mining-automation suppliers such as Caterpillar's own autonomy division and Komatsu have already deployed at scale across multiple sites globally.

The Diversification Signal, Read Skeptically

Procurement teams evaluating a foundation-model perception vendor for a non-automotive application should read this announcement as evidence that Helm.ai believes its technology transfers, not as proof that it already has. The US$70 million figure is cumulative across a trailing twelve-month period and across automotive, Tier 1 supplier, and industrial customers combined, with no breakdown of how much of that total came from the newer industrial and robotics segment specifically. A company expanding its addressable market by pointing to early industrial contracts, without disclosing what share of revenue those contracts represent, is a common and reasonable step in a diversification strategy, but it is not the same as demonstrating that the industrial business line has reached the scale the automotive business already has. Buyers in mining, warehousing, or other heavy-industrial segments evaluating Helm.ai alongside more established automation incumbents should ask directly for a reference site and a performance comparison against the incumbent they are currently running, rather than accepting the diversification narrative at face value.

There is also a credibility question specific to Helm.ai that the broader autonomous-vehicle software sector has had to reckon with more than once in 2026: the gap between a company's own reported financial figures and what independent trackers can verify from the outside. Helm.ai is privately held, and its total funds raised to date has been a point of disagreement between the company's own statements and third-party funding databases, a reminder that none of the figures in this announcement, the US$70 million contract total included, are independently audited. That does not mean the figure is wrong, but it does mean a buyer's diligence should treat it the way any unaudited, self-reported revenue claim from a private company deserves to be treated, as a starting point for verification rather than a settled fact.

What "Heading Toward Breakeven" Actually Signals

Helm.ai's statement that it is heading toward operating breakeven is the other detail worth parsing carefully, because it is doing more rhetorical work than it might first appear. A perception-software company moving toward breakeven on the strength of expanding commercial contracts, rather than continuing to rely on new funding rounds to cover losses, is a meaningfully different business profile than one still burning venture capital with no clear line of sight to sustainable revenue. If accurate, it suggests Helm.ai's foundation-model licensing business has reached enough scale across its combined automotive and industrial customer base to cover its operating costs, which would be a notable milestone in a sector where many AI-heavy perception and foundation-model companies remain years away from that point. The company did not disclose a specific timeline or a current burn rate alongside the claim, which means it should be treated as a direction of travel rather than a confirmed destination.

For buyers specifically evaluating foundation-model perception vendors for heavy equipment automation outside of mining, construction, ports, and agriculture are the adjacent categories most likely to see Helm.ai or a comparable vendor pitch similar unsupervised-learning advantages, since all three share mining's combination of harsh, variable outdoor conditions and limited existing labeled-data infrastructure. The practical next step for any buyer in those categories is not to wait for Helm.ai's next funding or contract announcement, but to request a proof-of-concept deployment scoped to the buyer's own site conditions, since dust, terrain, and lighting vary enough across sites that a system proven at one mining operation is not automatically proven at another. The contract total disclosed this week is a signal that Helm.ai believes the industrial pivot is working. Whether it is working well enough to displace an incumbent automation supplier at a specific site is a question only a direct, site-specific evaluation can answer.

Where Helm.ai Sits Against the Rest of the Perception Stack

Helm.ai is not the only company selling foundation-model perception software into automotive and, increasingly, industrial customers, and buyers should place this announcement against a competitive field that includes much larger, better-capitalized rivals. NVIDIA's DRIVE and Isaac platforms already serve automotive and industrial robotics customers with a far larger compute and software ecosystem behind them, while Mobileye has decades of automotive perception deployment history and existing relationships with the same Tier 1 suppliers Helm.ai is courting. Smaller specialists such as Wayve and Waabi have pursued their own end-to-end learning approaches to driving and perception, each making a version of the same generalization argument Helm.ai is making here. Helm.ai's distinguishing claim, that its unsupervised Deep Teaching approach needs less labeled data than rivals' systems, is a real technical differentiator if it holds up under independent scrutiny, but it is a claim every one of its competitors would dispute if asked, and none of them have published a head-to-head benchmark that would let a buyer judge the claim directly.

The scale of the announcement also deserves context against the capital flowing into this sector more broadly. US$70 million in cumulative commercial contracts over twelve months is a meaningful revenue figure for a private AI software company, but it is a fraction of the capital some automotive and robotics perception rivals have raised in single funding rounds this year, let alone the multi-billion-dollar compute commitments the largest humanoid and autonomous-vehicle programs have announced. That disparity does not necessarily put Helm.ai at a disadvantage if its software requires less compute and less labeled data to train and deploy, which is the entire premise of its technical pitch, but it does mean the company is competing against rivals with substantially deeper balance sheets if the market shifts toward a scale-and-compute arms race rather than a data-efficiency one. Buyers should ask directly how Helm.ai's compute and data requirements compare against the specific incumbent system they are currently running, since that comparison, not the funding totals on either side, is what will determine total cost of ownership at deployment scale.

The industrial and robotics expansion also raises a practical integration question that automotive customers have not historically had to deal with in the same way: whose safety certification framework governs a perception system deployed on a piece of mining or construction equipment that was not originally designed with autonomous operation in mind. Automotive perception software integrates into a vehicle platform with established functional-safety standards built around decades of regulatory precedent. Mining and construction equipment automation is governed by a less uniform, often site-specific and equipment-manufacturer-specific set of safety requirements, which means a buyer adopting Helm.ai's perception stack on a haul truck or excavator will likely need to work through a certification and validation process closer to a custom engineering project than a standard software procurement. That integration cost belongs in any total cost of ownership comparison a buyer runs against incumbent automation suppliers who have already been through that certification process on the buyer's specific equipment class.

This analysis synthesizes company statements and public market activity as of the publication date and should not be read as investment, financial, or professional advice; it is provided for general information purposes only.