AI in Defense: How Software Is Reshaping Weapons12 min read2,490 words

AI in Defense: How Software Is Reshaping Weapons

AI in defense is moving beyond prototypes. See how Tiberius Aerospace uses software, simulation and distributed manufacturing—and where the risks remain.

AI in defensedefense AIAI weapons manufacturingmilitary AI ethicsdefense technology startups
AI in Defense: How Software Is Reshaping Weapons

AI in Defense Is Becoming a Manufacturing Strategy

AI in defense is usually talked about in terms of autonomy - who decides, and what the system does on the battlefield. Tiberius Aerospace is framing it differently. Instead of putting software in charge of firing decisions, it's using software to analyze a military problem, simulate designs, and coordinate how equipment could be made. That shift matters: AI doesn't have to "pick" the final option to influence what gets built, how fast production moves, and which manufacturers get pulled in.

Tiberius was founded by former tech executives Chad Steelberg and Andy Baynes. WIRED reports that Steelberg started with an AI model to assess ways to affect the war in Ukraine. Tiberius says that work fed into Sceptre, a longer-range artillery munition, and into GRAIL, a software platform aimed at simulation, design iteration, and manufacturing coordination. The company has also announced major government and industry relationships, along with $85 million in Series A financing.

This looks less like a proof that software alone can solve battlefield or procurement problems - and more like a case study in how Silicon Valley-style tooling runs up against defense realities. The real questions are what AI changes in practice, how the business makes money, and what controls are needed when software helps shape lethal systems.

What problem does Tiberius say AI can solve?

Tiberius started with a big question: where could a new company actually make a difference in the war in Ukraine? Steelberg - then CEO of enterprise AI company Veritone - built an AI model to help assess that question. The model was meant to compare areas of work, not to control a weapon or make battlefield decisions.

According to the company, its analysis steered it away from improving defensive technology. The argument is that stronger defenses could reinforce a stalemate rather than change the balance. From there, the team looked at offensive systems, weighing things like cost-effectiveness, and tried to find where a new design could fill a capability gap.

That's how Tiberius ended up focusing on artillery ammunition instead of another drone. Its Sceptre projectile is designed for use with a NATO-standard 155-millimeter howitzer. The company says it's intended to travel farther than existing rounds and reports a successful test firing. Those are company claims about a developing system, not independent confirmation of battlefield performance.

Underneath the specific munition, the business problem is broader: can software help teams narrow a big, uncertain design space before they spend substantial time and manufacturing capacity? In principle, simulation can compare options under different conditions and make early-stage analysis faster and more repeatable - especially when engineers are dealing with lots of variables and constraints.

But simulation only works as well as its assumptions and data. A model can produce outputs that look persuasive while still being wrong or incomplete. Real effectiveness still depends on engineering, testing, procurement requirements, and feedback from the field. Tiberius says it's seeking active combat data, alongside open-source information. In practice, what it can credibly "tell" users will depend on whether that data is relevant, complete enough, and good enough.

This distinction matters for anyone evaluating AI companies. The more useful question isn't whether a business says it uses AI. It's what decision the system actually improves, what evidence supports that improvement, and what still has to be verified by humans and by physical testing.

How does AI fit into the weapons business?

Tiberius talks about two connected efforts, but they aren't the same thing. One is Sceptre, an artillery munition under development. The other is GRAIL - Generative Real-Time Artificial Intelligence for Lethality - a software platform the company says can simulate designs across different targets and environmental conditions. The stated goal is to help users test, iterate, and tailor designs more quickly.

It's worth keeping those separate. A software platform that supports design simulation isn't the same as a fielded weapon. And it isn't evidence that an AI system independently selects targets or authorizes attacks. The company's description focuses on design analysis, simulation, supply-chain management, and compliance. Any serious assessment has to start there: what those functions do, and what proof exists that they work the way the company says.

For a defense company, the software's value might extend beyond the first product. If GRAIL helps teams compare designs, manage documentation, and coordinate production across partners, it could become a kind of infrastructure that supports multiple development and manufacturing phases. That's a more durable business model than selling or licensing something tied to a single munition.

Tiberius has also announced work with the US Army to mature Sceptre, and a $500 million technology collaboration contract with UK partners to manufacture, license, and distribute the munition. The company has said it's developing a missile it expects to start generating revenue from the following year. These announcements suggest real interest - but they don't automatically mean a system is operational, profitable, or available at scale.

One practical way to think about AI's role is to break it into four stages:

  • Problem selection: Does the analysis point to a real capability gap, or does it just optimize for a convenient metric?
  • Design exploration: Do the simulations actually help narrow options, and are the assumptions documented?
  • Validation: Are promising outputs tested through the right engineering and procurement steps?
  • Deployment and oversight: Who approves changes, controls access, and stays accountable for how the system is used?

That framework applies beyond defense. In any high-stakes industry, AI's value isn't just whether it can produce an answer - it's whether it plugs into a controlled decision process.

Why does manufacturing matter as much as design?

A good design doesn't matter much if you can't manufacture it at meaningful scale, deliver it reliably, and keep it aligned with procurement and regulatory requirements. That's why Tiberius's manufacturing pitch is central. Baynes describes an approach influenced by Apple: keep the core design and coordination capabilities inside the company, while relying on a network of external manufacturers and suppliers.

The advantage, in theory, is flexibility. A distributed manufacturing network could tap spare precision-production capacity across allied countries, without the cost of owning every facility. It may also make it easier to expand production without building a huge fixed-cost manufacturing base during peacetime. Tiberius says GRAIL is intended to support that by automating parts of supply-chain management and compliance, along with design simulation.

The trade-off is control. Outsourcing can lower capital needs and speed up access to capacity, but it complicates quality assurance, supplier security, traceability, and coordination. Defense production also comes with requirements that don't show up in the same way in consumer electronics. A design change that's easy to "push" in software may still require review, testing, approvals, and production updates before it's safe and legal.

That's where the Apple analogy has limits. Consumer products often get frequent software updates, fast. Physical defense systems don't. They come with longer validation cycles, safety implications, and procurement obligations. Iteration can be helpful, but speed isn't the only goal. Teams still have to maintain configuration control: they need to know which version was tested, which parts were produced, and what evidence supports each change.

And a distributed model depends on trusted partners. Manufacturing capacity isn't interchangeable just because machines exist. Suppliers need the right expertise, secure processes, and clear contractual responsibility. When multiple companies touch production, the prime contractor still has to coordinate standards and accountability in a coherent way.

For other businesses, the lesson isn't simply "outsource more." It's to figure out which capabilities create durable advantage, and which can realistically be accessed through partners. In Tiberius's framing, software, design coordination, and access to a manufacturing ecosystem reinforce each other. Whether that holds up at scale comes down to execution: partner reliability, quality controls, compliance, and the ability to deliver consistently.

What can other AI businesses learn from this case?

Start with a constrained operational problem, not a technology searching for a use case. Tiberius's initial question was broad, but its process tried to narrow the field by comparing potential interventions. That's a useful pattern for AI projects anywhere: define the decision or bottleneck, identify what evidence exists, then test whether software actually changes outcomes.

Second, a model doesn't always have to be the final product customers pay for. Here, the early AI work helped the founders evaluate an opportunity, and the company later built a product concept around simulation and manufacturing coordination. In other sectors, an internal analysis tool can turn into a repeatable workflow that's more commercially useful than the original model itself.

Third, design around the surrounding system. Even if an AI tool generates a promising simulation, you still need users, data, testing processes, procurement pathways, and production capacity. Many teams underestimate those integration costs. In regulated markets, documenting and governing decisions can matter as much as raw model performance.

There's also the financing and customer-acquisition reality. Defense tends to mean long timelines, demanding buyers, and high barriers to entry. Tiberius's announced financing and partnerships may help it reach testing and manufacturing milestones, but announcements aren't the same as shipped products, repeat customers, or demonstrated performance. A founder looking at a similar market should treat interest, collaboration agreements, qualification, procurement, and revenue as separate stages with different risks.

Here's a practical checklist for AI businesses operating in complex industries:

  • Name the workflow: Which decision or process is changing?
  • Set a baseline: How is that work done today, and what does it cost in time, money, or risk?
  • Specify evidence: What data is needed, and how will its quality and limits be checked?
  • Keep humans accountable: Who reviews outputs and has authority to reject them?
  • Plan for integration: Which existing systems, regulations, and partners must be involved?
  • Measure delivery, not announcements: What has been tested, approved, deployed, and paid for?

It's easy to overlook that last point in an emerging market. A flashy strategy can attract capital. The durable business is built through repeatable delivery. The more consequential the application, the more important it is to show where the system has been tested - and where it still can't be relied on.

What are the ethical and governance limits?

AI in weapons raises questions that differ from many commercial uses - not just in degree, but in kind. A design tool can influence what capabilities get developed, how quickly they're iterated, and where production is directed. Even if it never makes targeting decisions itself, it can still shape the menu of options available to military and government users.

That makes transparency about function a big deal. Companies should distinguish between decision support, design simulation, logistics automation, and operational control. Those categories carry different risks and call for different safeguards. In the material reported about Tiberius, GRAIL is described as a platform for simulation, design iteration, supply-chain management, and compliance. That description shouldn't be stretched into claims about autonomous targeting - or treated as a complete picture of how the system will be used.

Data governance is another concern. Combat data may be sensitive, incomplete, or shaped by the conditions under which it was collected. If a simulation is trained or calibrated on limited evidence, it can give users an illusion of precision. Users need to understand data provenance, uncertainty, and the scenarios the system doesn't represent. A useful model should surface its assumptions instead of hiding them behind a polished interface.

There's also accountability across a distributed supply chain. If software recommends a design change, a manufacturer builds a component, and a government customer deploys the finished system, responsibilities can get fragmented fast. Contracts and technical processes should make clear who validates changes, who controls access to sensitive data, and who's responsible when performance doesn't match expectations.

Finally, the business model deserves scrutiny. Steelberg has said the company expects to make money, alongside describing a desire to affect the war's outcome. Commercial incentives don't automatically invalidate a defense product, but they do make strong governance and independent evaluation essential. Claims about effectiveness should be separated from fundraising, valuation, and partnership announcements.

Key takeaways:

  • AI can shape defense capability upstream, through analysis and design, without directly controlling a weapon.
  • Simulation is only useful when assumptions, data limits, and validation steps are clear.
  • Distributed manufacturing can add capacity, but it raises the stakes on supplier security and quality control.
  • Commercial traction should be judged by tested systems and accountable deployment - not just headline numbers.
  • In high-consequence settings, faster iteration should not mean weaker review.

Conclusion

Tiberius's story isn't just about applying AI to weapons. It's about using software to connect problem analysis, design simulation, and a distributed manufacturing strategy in a sector where delivery is hard and consequences are high. That approach may help teams explore options faster, but it doesn't remove the need for testing, procurement discipline, or accountability.

For businesses considering AI in similarly complex markets, the useful question isn't whether a model can generate an answer. It's whether the full system - from data and validation to partners and governance - can turn that answer into something reliable and responsible.

Frequently Asked Questions

Tiberius Aerospace is a defense technology company founded by Chad Steelberg and joined by former Apple executive Andy Baynes. It is developing the Sceptre artillery munition and the GRAIL software platform, which the company says supports simulation, design iteration and manufacturing coordination.
The company says it uses AI and simulation to analyze potential design options and assess how systems may perform under different conditions. GRAIL is also intended to support supply-chain management and compliance. The reported work describes design and operational support, not an AI system independently authorizing attacks.
Sceptre is a Tiberius-developed projectile designed to be fired from a NATO-standard 155-millimeter howitzer. The company says it is intended to travel farther than currently available rounds and has announced a successful test firing. Those claims do not, by themselves, establish operational performance or deployment at scale.
The company says it wants to use manufacturing capacity across partner networks rather than own all production facilities. That could reduce fixed costs and provide access to more capacity, but it also creates challenges involving supplier security, quality assurance, compliance and coordination.

Related articles