Enterprise AI costs are unsustainable, says ManageEngine

• ManageEngine says enterprise AI costs are already unsustainable, and it is betting on efficiency over frontier-scale power.
• CEO Rajesh Ganesan ties the bet to a China strategy built around Beijing’s push for domestic tech.

Enterprise AI costs have become the reckoning that the industry has put off. After two years spent raising hundreds of billions of dollars on the promise of ever more powerful models, the harder question of who pays to run them, and whether the economics hold, is only now being asked out loud. Rajesh Ganesan has been asking for it from the start. “Does everyone need a frontier model? I don’t think so,” the ManageEngine CEO says, and coming from a company that has stayed private and profitable for two decades, the claim carries a weight the frontier labs’ projections do not.

“We don’t throw the kitchen sink at problems,” he said, framing efficiency as a design principle his company has held from the start rather than a reaction to spiralling bills.

The case that enterprise AI costs are already breaking

The sharpest version of that case is about consumption. Ganesan argues the token-based model of buying AI is already unsustainable, and he uses ManageEngine’s own developers to make the point. When AI-assisted coding was limited to experienced engineers, he said, token costs stayed under control because a senior developer knows a ten-line fix only needs ten lines sent to the model.

Once the tools reached a few thousand employees, younger engineers were sending a hundred files to do the same job, simply because they did not know better. Extend that behaviour to a sales team attaching spreadsheets and research documents to every query, and the cost curve stops looking like a rounding error.

Ganesan’s preferred analogy is aviation. Supersonic passenger travel was retired in 2003, not because it was slow, but because it was never economically sustainable, while subsonic jets went on to move far more people at a fraction of the cost. He applies the same logic to AI. Chinese labs, he notes, have already shown that competitive models can run on smaller compute, which raises an awkward question for anyone budgeting around frontier-scale infrastructure.

The company’s position is that it would rather build efficient models than price for expensive ones. “Only when we force the models to be disciplined [only then] can we be in a position not to price it,” Ganesan said.

That framing is not new for the company, and enterprise buyers should read it as a strategy as much as a philosophy.When ManageEngine rolled out its autonomous Zia Agents in May, Ganesan was already on record arguing that frontier models are strong for general use but poorly suited, on cost, to the narrower demands of enterprise IT. The Jakarta sessions were a recap, pulling that worldview together with everything the company has shipped this year.

The autonomous Zia Agents rollout on May 21 extended AI from assistance into autonomous execution across the suite, from service desks through to security operations. Nine days earlier, ManageEngine had rebuilt its Log360 security platform around native SOAR, and by the end of June, it had opened a developer marketplace that already lists more than 170 partner-built extensions.

Two threads run through all of it: agents the company says are never trained on customer data, and support for the model context protocol so customers can plug in third-party models. Both feed the same pitch, that governance is easier when one vendor controls the stack.

Efficiency meets sovereignty in China

Where the efficiency argument gets its real-world test is China. ManageEngine has operated there for roughly two decades, and Ganesan is candid that the market runs on different rules. Beijing’s Xinchuang policy, formalised in a 2022 directive known as Document 79, requires state-owned enterprises and critical sectors to replace foreign software with domestic alternatives by 2027.

Rather than treat that as a wall, the company is building for it. A team of around 250 people is reworking its products with local components so they qualify as sufficiently Chinese, and its hosted deployments there already run on Alibaba’s Qwen models.

The mechanism that makes this workable is a gateway the company calls Platform AI, which lets customers choose which model sits underneath. The default is Zia’s own model, with options including Claude and Qwen, so a bank in a regulated market can decide who runs its data without leaving the product.

It is a tidy expression of the same instinct that drives the cost argument. Keep the expensive, opinionated parts optional, and let the customer control the bill and the jurisdiction. China also shows how the company sells. Ganesan said between 70 and 80% of its China business goes through partners, with direct engagement reserved for customers who insist on it, a channel-heavy structure that mirrors how it operates across most of the world.

A small slice, growing fast

Southeast Asia, for all the profile of a user conference held in the region, is not yet where ManageEngine makes most of its money. It counts more than 7,500 customers across key ASEAN markets, with over 800 in Malaysia and names such as AirAsia and CIMB among its regional references, but the region remains a small part of the whole.

Arun Kumar, ManageEngine’s regional vice president, said Southeast Asia has grown at around 25% a year for the past four or five years, and still accounts for less than 10% of group revenue. The company is private and does not disclose absolute figures.

The bigger shift has been away from the United States. US revenue has fallen from around 70 to 75% of the group total to between 35 and 40%, Kumar said, the result of a deliberate move over the past eight to ten years to open offices across more markets and avoid leaning too heavily on any single country.

Southeast Asia is one beneficiary of that spread, growing fast off a small base rather than carrying the company. What gives the region its own character is how much of it still runs on-premises. Roughly half of ManageEngine’s Southeast Asia revenue sits on-prem, Kumar confirmed, weighted toward banking and finance customers that are slower to move workloads, and slower to hand their data to someone else’s cloud.

It is a customer profile that rewards exactly the cost-controlled, sovereignty-aware pitch Ganesan is making. None of this is disinterested. ManageEngine does not build frontier models, so an argument that enterprises do not need them is also an argument for what it happens to sell.

A frontier lab would counter that today’s efficient model is last year’s frontier commoditised, and that the ceiling keeps moving. Ganesan is comfortable with that trade. He bets that once the invoices arrive, most enterprises will realise they paid frontier prices for work that never needed the frontier.”