New From Our Reporters
Exclusive
The Startup Trying to Tame Accounting Chaos Behind AI Data Centers
The rush to build and finance AI data centers has created a new kind of stress for CFOs at companies developing them.
Some executives at these firms tell me they are struggling to keep their books accurate and move money from their lenders to their suppliers on time, raising the risk of costly delays or even huge overpayments.
That risk is creating an opening Kos.AI, which is developing software that functions like a virtual employee, reviewing dense invoices and contracts.
The San Francisco-based startup recently raised $12 million in a round co-led by 8VC and XYZ Ventures.
The round included angel investors with deep ties to the industry, including Sachin Katti, OpenAI’s head of compute infrastructure; former Crusoe CFO Matthew DeNezza; and longtime Equinix CFO Keith Taylor. (Kos CEO Tanuj Thapliyal and his cofounder and CTO, Manikanta Kotaru, were previously graduate researchers in Katti’s Stanford University lab focused on large-scale intelligent systems.)
As companies prepare to pour trillions of dollars into data centers, power plants and other physical infrastructure, the back-office work for processing that spending is becoming overwhelming, people in the industry tell me.
Reviewing invoices, tracking purchase orders and requesting funds from lenders is time-consuming, eating into finance teams’ nights and weekends.
Thapliyal told me that a single invoice tied to a data center project can run more than 800 pages, often accompanied by contracts that can stretch to thousands of pages.
Much of the work of reviewing them is still handled manually, and accountants with construction finance expertise are in short supply.
"There are not enough trained people to be able to review all of this, and these workflows are entirely manual,” Thapliyal said.
He said a typical data center developer may receive monthly invoices from a general contractor totaling more than $500 million.
Finance teams then dive into contracts to verify every detail, checking labor hours, equipment pricing and whether orders match what was delivered.
Executives tell me most finance teams simply weren’t built for this level of volume or velocity.
The worst-case scenario, one data center CFO said, is getting a call from a general contractor about a missing or late payment and scrambling to track it down.
Some teams are resorting to a risky workaround: focusing only on the largest line items, such as equipment purchases, and hoping they didn’t miss any discrepancies or overcharges buried in the details, according to a finance executive at a large data center developer.
Thapliyal says this is exactly the kind of work AI should handle.
Kos’s system connects to a company’s email, Slack, Zoom and core applications such as Excel and internal enterprise resource planning systems.
When invoices come in, for instance, employees can forward them to Kos via email for review.
Employees can ask Kos questions about the contracts or ask for status updates.
The company, which was founded about six months ago, says it has signed multiple customer contracts.
Because Kos aims to act like a virtual employee, it costs roughly the same amount of money as a human worker, Thapliyal said, so most contracts are priced in the six-figure range.
Unlike a human, though, Kos can work 24/7, he said. (The founders say the name “Kos” draws on Japanese roots tied to the idea of personalization, which is how they view software evolving.)
To make sure Kos doesn’t make mistakes when dealing with a company’s financial operations, the firm developed a method where Kos’ work must be rooted in a company’s own well-defined processes.
Still, Thapliyal said each of his customers pair the AI “employee” with a human manager, adding a layer of oversight.
Before pursuing an MBA at Stanford, Thapliyal worked at Meraki, a cloud networking firm.
Meraki CEO Sanjit Biswas, who later founded Samsara, introduced Thapliyal to Katti.
At Stanford, Thapliyal and Katti co-authored a research paper on identifying software bugs that caused applications to consume excessive data center power.
While Kos is focused on selling to AI infrastructure firms, it is also targeting other capital-intensive industries like energy and defense.
That helps explain its choice of VC backers; 8VC and XYZ have invested heavily in defense startups such as Anduril.
I haven’t come across too many AI startups developing apps for the fast-growing infrastructure sector.
But I get the sense that people building AI infrastructure or financing it would be a ripe target for this kind of product, as the status quo definitely isn’t cutting it.
In Other News….
Google is in talks with Marvell Technology to develop two new chips aimed at running AI models more efficiently, https://www.theinformation.com/articles/google-talks-marvell-build-new-ai-chips-inference">my colleague Qianer reported
Business Insider https://www.theinformation.com/briefings/xai-rent-computing-power-cursor">reported that SpaceX subsidiary xAI will rent computing power from its data centers to the coding startup Cursor.
OpenAI agreed to pay chip designer Cerebras more than $20 billion to use servers powered by the firm’s AI server chips over the next three years, https://www.theinformation.com/articles/openai-spend-20-billion-cerebras-chips-receive-equity-stake">The Information reported.