# Is the Space Industry Ready to Let AI Run Its Satellites?

**There are approximately 16,000 active satellites in Earth orbit today — and the industry is structurally unprepared to operate what comes next.** Market intelligence firm Novaspace projects up to 43,000 satellites will be built and launched over the coming decade. Goldman Sachs puts an even more aggressive number on the [Low Earth Orbit (LEO)](https://orbital-intel.com/glossary/leo) build-out: as many as 70,000 new LEO satellites launched within five years alone. At that scale, traditional ground-based human operations teams are not a bottleneck — they are a ceiling. The argument for onboard AI autonomy is no longer theoretical; it is an operational necessity driven by sheer constellation math.

That is the core argument advanced by Martin Halliwell, a Partner at NewSpace Capital — described in the source as one of the world's first private equity firms devoted exclusively to growth-stage space technology companies — and former Chief Technology Officer of SES, where he led global technology and R&D from 2011 to 2019. Writing in Space.com's Expert Voices section, Halliwell makes a systematic case for why AI must move from advisory role to decision-maker aboard satellites, while being candid about the legal, organizational, and trust barriers that will slow adoption.

This is a perspective worth taking seriously. Halliwell spent nearly a decade running the technical architecture of one of the world's largest [satellite constellation](https://orbital-intel.com/glossary/constellation) operators, and he now sits on the investment side evaluating which companies will be positioned to solve this problem commercially.

---

## Why Human Operations Don't Scale Past a Certain Constellation Size

The core physics of the problem is straightforward. A network of a few satellites can be commanded from the ground with tolerable latency. A network of hundreds or thousands cannot. As Halliwell notes, some decisions aboard a satellite must be made in seconds — collision avoidance responses, beam re-pointing during a demand surge, anomaly handling — without waiting for a round-trip command from a ground station.

The specific operational use cases he identifies:

- **Network capacity management:** Demand for satellite bandwidth is not static. It spikes over cities at peak hours, around major events, in disaster zones, on aircraft and ships, and during military operations. An AI system tracking these changes could dynamically direct where a satellite's beams point, how much power each beam draws, and when capacity should shift. The efficiency gain is real: instead of fixed allocations to fixed footprints, spectrum and power — both finite aboard any spacecraft — are directed where they are actually needed.

- **Onboard data processing for Earth observation:** Currently, most satellite imagery and sensor data is downlinked to Earth, then cleaned, sorted, and analyzed on the ground before anything useful reaches a customer. Halliwell argues that moving more of this pipeline into orbit — processing data onboard and downlinking only actionable results — cuts costs, reduces latency, and in defense applications could meaningfully accelerate decision cycles. It also reduces dependency on ground infrastructure, which can itself be targeted or disrupted.

The logic for autonomous systems here is not exotic. It mirrors what aviation autopilots, algorithmic trading systems, and industrial process controllers have done in other high-tempo, high-complexity domains for decades. The satellite industry is simply arriving later.

---

## The Non-Technical Barriers Are the Real Bottleneck

Halliwell's most valuable contribution in this piece is the frank accounting of where the actual friction lives — and it is not in the AI models themselves.

**Legal ambiguity is the immediate problem.** Most operating licenses, insurance policies, and contracts governing satellite operations were written under the assumption that a named human operator makes each consequential decision. When an AI system autonomously re-points a beam and disrupts an adjacent service, or makes a maneuver decision that leads to a conjunction, the question of liability — operator, manufacturer, or software vendor — becomes genuinely murky. No current regulatory framework cleanly resolves this.

**Operator risk tolerance is conservative by necessity.** Halliwell observes that many operators are comfortable with AI in an advisory capacity. Fewer are willing to grant it authority to reconfigure a network or shift capacity without explicit human approval. This is rational given the liability picture above, not mere technophobia. The practical result is that adoption will be incremental: operators will define carefully bounded decision domains where AI can act autonomously, and maintain human-in-the-loop requirements everywhere else.

**Data security and model governance are infrastructure problems.** To train and deploy AI models on sensitive satellite operational or imagery data, operators need secure, compartmented environments with strong access controls and clear data-separation guarantees. In some cases — particularly defense — models may need to be trained entirely within classified or restricted networks, so sensitive data never transits an external system. Halliwell frames this not just as a technical challenge but a trust challenge, one that requires standards, auditing, and demonstrated reliability over time.

---

## AI in the Design and Manufacturing Loop

Beyond operations, Halliwell identifies a second near-term impact: AI acceleration of spacecraft design and manufacturing. Engineers using AI to draft code, search technical documentation, generate preliminary structures, antenna configurations, power system layouts, or mission plans could compress development timelines and reduce the team sizes needed for a given scope of work. AI-assisted factory quality control — spotting manufacturing defects, predicting schedule delays, flagging equipment maintenance needs — fits the same pattern.

The framing here is appropriately measured. AI handles routine work faster; humans still own the difficult judgment calls, safety analysis, and the hardest technical problems. The value is in shifting the ratio of time engineers spend on high-leverage versus low-leverage work, not in replacing engineering judgment.

For investors evaluating [NewSpace](https://orbital-intel.com/glossary/newspace) companies, this is relevant context: startups that can credibly demonstrate shorter development cycles through AI-assisted engineering carry a real cost and time-to-market advantage, particularly in a capital environment where runway conservation matters.

---

## Industry Trajectory: Skeptical Read

The satellite autonomy thesis is directionally correct, but the timeline to meaningful autonomous authority is almost certainly longer than AI optimists project. Halliwell himself signals this with "adoption is likely to be gradual."

A few structural realities worth noting that the op-ed does not fully surface:

**Regulatory reform is slow.** ITU coordination, national licensing regimes, and insurance underwriting standards all move on multi-year cycles. Even if operators want to grant broader autonomous authority to their satellites, the regulatory scaffolding to do so legally and insurably may not exist on the same timeline as the technical capability.

**The megaconstellation operators are already ahead of the rest of the market.** [SpaceX](https://orbital-intel.com/companies/spacex) has been running autonomous collision avoidance across Starlink for years. The autonomy problem is most acute for the second and third tier of constellation operators — the regional broadband providers, the Earth observation players, the IoT networks — who do not have the same internal engineering depth to build proprietary autonomy stacks. This is where third-party AI operations software has the clearest commercial opening.

**Ground systems are not going away.** Moving some analysis into orbit reduces ground dependency at the margin; it does not eliminate it. The architecture is more likely to evolve toward edge-cloud hybrid models — onboard AI handling time-critical decisions, ground systems handling everything that can tolerate latency — rather than a wholesale shift to orbital autonomy.

For coverage of autonomous robotic systems applied to on-orbit operations and assembly — an adjacent problem set — [humanoidintel.ai](https://humanoidintel.ai) tracks that sector in detail.

---

## Key Takeaways

- **16,000 active satellites today; Goldman Sachs projects up to 70,000 new LEO satellites in five years** — human operations teams cannot scale to match.
- **Onboard AI for capacity management and Earth-observation processing** are the two most commercially mature near-term use cases identified by former SES CTO Martin Halliwell.
- **Legal liability, operator risk tolerance, and data governance** are the real blockers — not the AI technology itself.
- **Adoption will be incremental:** operators will carve bounded autonomous decision domains while maintaining human approval for higher-stakes actions.
- **Third-party AI operations software** has the clearest commercial opening among mid-tier constellation operators who lack the internal resources to build autonomy in-house.
- **Regulatory frameworks** governing satellite licensing and insurance were written for human-in-the-loop systems and will require reform before full autonomous authority is legally and insurably viable.

---

## Frequently Asked Questions

**Why can't human operators just scale up staffing to manage larger constellations?**
The economics and physics don't support it. At constellation sizes of hundreds to thousands of satellites, the volume of telemetry, conjunction assessments, beam management decisions, and anomaly responses exceeds what any realistically sized human team can process in real time. Some decisions — particularly collision avoidance — must be executed in seconds, faster than any ground-loop command cycle allows.

**What specific satellite functions is AI best suited to handle autonomously?**
According to Halliwell's analysis, the strongest near-term cases are dynamic bandwidth and beam management (tracking demand shifts and directing capacity accordingly) and onboard data processing for Earth observation (filtering raw imagery or sensor data in orbit before downlink). Both involve high-volume, pattern-recognition tasks well suited to current AI architectures.

**Who is liable when an AI system on a satellite causes interference or damage?**
This is currently unresolved. Existing satellite operating licenses, insurance policies, and contracts generally assume a human decision-maker is responsible for each consequential action. When an autonomous AI system makes a decision that causes harm, responsibility across the operator, spacecraft manufacturer, and software provider is legally ambiguous — and no major regulatory framework has cleanly addressed it yet.

**How does onboard AI processing improve defense satellite utility?**
By moving data analysis closer to the sensor — processing imagery or signals intelligence onboard and transmitting only actionable results — latency from collection to decision is reduced. It also reduces dependence on potentially vulnerable ground processing infrastructure, making the overall system more resilient to disruption.

**What does this mean for companies building satellite operations software?**
The commercial opening is largest among mid-tier and smaller constellation operators who lack the internal engineering depth to build proprietary autonomy stacks. Third-party AI-driven operations platforms that can demonstrably handle bounded autonomous decisions — with auditable logs, clear human-override mechanisms, and a credible liability framework — are well positioned as constellation counts continue climbing.