The Metered Moat

In October 2023, the Wall Street Journal reported that GitHub Copilot was costing Microsoft more than $20 per user per month, and as much as $80 for heavy users, while the product sold for $10. The most sophisticated software company on earth was paying customers to use its flagship AI product, and it was doing so on purpose. That single number quietly rewrote the oldest strategy question in SaaS: what matters more, network effects or margin?

For twenty years the answer was easy, classic software carried 80 to 90 percent gross margins because serving the next customer cost almost nothing, so margin took care of itself while distribution and lock-in stayed scarce. You ran the participation layer thin or free, the way Slack and Figma did, because the marginal user was nearly costless and the network those users formed was the moat. Margin was a bookkeeping outcome rather than a strategic choice.

AI broke both halves of that logic at the same time.

The meter is now running on every user

Inference is a real cost that recurs on every query, which means the marginal user is no longer free. Bessemer Venture Partners put it plainly in its February 2026 pricing playbook: "Every AI query incurs real compute costs. Companies see 50-60% gross margins vs. 80-90% for SaaS." ICONIQ's survey data shows AI-native gross margins climbing from 41 percent in 2024 to roughly 52 percent in 2026, improving but nowhere near the classic software profile, and the fastest-growing cohort in Bessemer's benchmarks runs closer to 25 percent, often negative.

So the old move of giving the product away because the COGS rounds to zero now has a meter attached. You can still buy growth with a free tier, but you are buying it with real money, and you can bleed to death doing the exact thing that built the last generation of winners.

The obvious response is to wait for inference prices to fall, and they are falling fast. a16z calls the roughly tenfold annual decline for equivalent model performance "LLMflation." The catch, per the same analysis, is that token consumption is growing faster than per-token prices are dropping, a textbook case of Jevons Paradox, the nineteenth-century observation that making coal cheaper made Britain burn more of it. Cheaper inference relieves margin pressure per query while raising the total bill, so falling prices are a tailwind for the bet, not a substitute for making it correctly.

The same technology devalued every moat except the network

While AI was making users expensive, it was making software cheap to replicate. Any feature a product ships, a competitor with a frontier model API can approximate in months, and AI agents are collapsing the switching costs that workflow lock-in depended on. (Also read: What Survives When AI Becomes a Commodity)

Satya Nadella went further than most vendors would dare on the BG2 podcast in December 2024: "The notion that business applications exist, that's probably where they'll all collapse, right, in the agent era. Because if you think about it, they are essentially CRUD databases with a bunch of business logic." 

What survives that collapse lives outside the code. The graph of users who need each other and the proprietary usage data that makes the product better with every interaction. The cost of leaving a place where everyone you work with already is. Network effects went from one moat among several to nearly the only one left, at the precise moment they stopped being free to build.

Railroads already ran this experiment

The nineteenth century tested this exact configuration. Railroads were the ultimate network-effects business of their era, since every new station made every existing station more valuable, but unlike software, every mile of track cost real capital. During the British Railway Mania of the 1840s, investors funded thousands of miles of proposed track on the theory that any network would eventually pay, and a large share of it connected demand that did not exist. The lines carrying real traffic went on to become some of the most durable monopolies in industrial history, while the crash punished people who had paid network-building prices for things that were not networks.

That is the AI SaaS market in 2026, almost line for line.

The Metered Moat

Call the new configuration the Metered Moat where network effects remain the strongest defensibility available, perhaps the only durable one left, but the participation layer that builds them now has a running meter. Margin stops being a competing objective and becomes the constraint on how fast you can afford to buy the network, which means the question stopped being "network effects or margin" and became "what is the maximum sustainable burn rate on the participation layer, and is the network we are buying with it durable?"

One question separates a real bet from an expensive delusion. Does usage by one customer make the product better for other customers, through the graph, through shared data, or through interoperability? If yes, tolerate ugly margins on that layer as long as unit economics converge with scale, because you are buying the one asset a competitor cannot generate with an API key. If no, you do not have a network-effects business, and margin discipline is your entire strategy, because efficient execution is all you have. (Score your own product against the Network Effect Scorecard)

Both failure modes are already visible

The current market is producing casualties on each side of that test. One is the margin-first AI company, profitable-looking, gating every feature behind usage limits, building nothing a competitor cannot copy. The moment someone undercuts them, no graph holds the customers in place. The other is the growth-first company subsidizing inference for a "network" that is not one, lots of users producing no user-to-user value, a leaky bucket purchased at negative margin. A million people using your AI tool individually is not a network but a cost center with a churn problem, and subsidizing it at negative margin is the most expensive way to learn the difference.

So when the board asks the old question, reframe it. Margin functions as the purchasing power you spend on network effects rather than as their competitor, and the one unforgivable error left in this market is spending that power on a network that does not exist.

Key Takeaways

• AI inference resets SaaS gross margins from the classic 80 to 90 percent down to 50 to 60 percent, per Bessemer and ICONIQ data, so the marginal user carries a real cost for the first time in software history.
• The same models compressing margins are devaluing feature and workflow moats, leaving the user graph, proprietary data flywheels, and interoperability as the defensibilities that hold.
• The Metered Moat reframes the tradeoff: margin functions as the budget for buying network effects rather than a competing objective, so the strategic question becomes maximum sustainable burn on the participation layer.
• One test separates the bet from the delusion: usage by one customer must make the product better for other customers, or you do not have a network-effects business.
• A million individual users of an AI tool with no user-to-user value creation forms a cost center with a churn problem, and subsidizing it at negative margin is buying a leaky bucket.
• Falling inference prices help the bet survive but do not replace it, since total token consumption is growing faster than per-token prices are dropping.

FAQ

What is the Metered Moat?

The Metered Moat describes the AI-era condition where network effects remain the strongest available defensibility while the free-user layer that builds them carries real inference costs. The moat is still worth building, but a meter now runs while you build it, so margin becomes the constraint on how fast you can afford to buy the network.

Are network effects or margins more important for AI SaaS companies?

Network effects, with a condition. They are one of the few moats AI has not devalued, so a real network justifies tolerating compressed margins as long as unit economics converge with scale. If usage by one customer does not improve the product for others, no network exists, and margin discipline becomes the entire strategy.

Why are AI SaaS gross margins lower than traditional SaaS?

Every AI query spends real compute, so cost of goods scales with usage instead of amortizing across it. Bessemer's February 2026 data puts AI gross margins at 50 to 60 percent versus 80 to 90 percent for traditional software, and ICONIQ's survey average for AI-native companies sits near 52 percent in 2026.

About the Author

Martin Goetzinger has spent his career in enterprise software sales, helping large organizations such as Apple, Microsoft, and Verizon connect data, insight, and action. His work focuses on transforming how businesses measure success and create customer value through technology.

Outside the enterprise world, he writes about the five forces he believes are reshaping everything: AI, blockchain, energy, personalized health, and robotics. Not from a purely technical lens, but from a human one as to how these technologies will redefine work, wealth, and well-being.

He is based in the U.S. and publishes at www.MartinGoetzinger.com.

Disclaimer

The views expressed in this article are the personal opinions of the author and are provided for informational and educational purposes only. Nothing in this article constitutes investment advice, financial advice, legal advice, or any other form of professional advice. Do not make investment or financial decisions based on the content of this article. Always consult a qualified professional before making decisions that affect your finances, business, or livelihood.