The AI investing glossary

The terms readers ask about most

• What is a hyperscaler?

  The handful of companies that build and run data centers at global scale, and whose capex drives the AI build-out.

• What is a neocloud?

  An AI-focused cloud that rents GPU capacity, distinct from the traditional hyperscalers and usually financed with debt and long contracts.

• What is high-bandwidth memory (HBM)?

  Stacked DRAM placed next to an AI accelerator to feed it data fast enough; the reason memory became one of AI’s choke points.

• What are remaining performance obligations (RPO)?

  Revenue a company is contractually owed but hasn’t yet earned: the closest thing in the filings to a measure of contracted future demand.

• What is an AI data center?

  A data center built around accelerators: far more power per rack, liquid cooling, and campuses now measured in gigawatts.

• TPU vs GPU: what is a TPU?

  Google’s custom AI chip versus the general-purpose GPU, and why the hyperscalers’ custom silicon is the main long-term question for Nvidia.

• What is CoWoS? Advanced packaging for AI chips

  TSMC’s method of joining an AI processor and its HBM stacks on one package, and one of the tightest bottlenecks in AI chip supply.

• What is vendor financing?

  A seller helping its customer pay for what it buys; in AI it ranges from Nvidia’s lease guarantees to AMD’s customer warrants.

Every term, A to Z

AI accelerator

A chip designed to run AI calculations in parallel: GPUs from Nvidia and AMD, and custom chips such as Google’s TPUs and Amazon’s Trainium. Accelerators are the largest single cost in an AI data center. See the chips layer and TPU vs GPU.

AI data center

A data center built around accelerators: far more power per rack, liquid cooling, and campuses now measured in gigawatts. Read the full entry.

Backlog

A company’s own measure of orders received but not yet delivered. Unlike remaining performance obligations, backlog has no standard definition and may include orders that can be cancelled, so compare it only within one company over time.

Capacity buy-back

A supplier’s promise to buy unused capacity back from its customer. Nvidia discloses tens of billions of dollars of commitments to buy cloud capacity from AI cloud partners, which reduces their risk of idle GPUs. See the money-loop map.

Capex (capital expenditure)

Cash spent on long-lived assets such as data centers, servers and land, reported in the cash-flow statement as purchases of property and equipment. Our figures use cash capex and leave out finance leases, which some companies use heavily on top. See the capex tracker.

Circular financing

Suppliers investing in, lending to or guaranteeing their own customers, so that part of the demand they report is demand they helped pay for. Every link in the AI stack is on the money-loop map.

Cloud service provider (CSP)

A company that rents computing over the internet. Suppliers such as Nvidia use the term in their filings for both hyperscalers and smaller clouds. See hyperscaler and neocloud.

Convertible note

A loan that can turn into shares later, usually at the next funding round or an IPO. Amazon’s first $8 billion in Anthropic took this form; see who funds Anthropic.

CoWoS (advanced packaging)

TSMC’s method of joining an AI processor and its HBM stacks on one package, and one of the tightest bottlenecks in AI chip supply. Read the full entry.

Custom silicon (ASIC)

A chip designed for one customer’s workloads, such as Google’s TPU or Amazon’s Trainium, usually with a design partner like Broadcom or Marvell. Custom chips are the hyperscalers’ main way to reduce their dependence on Nvidia. See TPU vs GPU.

Depreciation and useful life

The cost of equipment spread over the years it is expected to be used. The longer the useful life a company assumes for its servers, the lower its yearly depreciation and the higher its reported profit. See the depreciation calculator.

Equity-method investment

An investment large enough to give significant influence, usually 20% to 50%, for which the investor books its share of the company’s profit or loss. Microsoft accounts for OpenAI this way, which is why it reports gains and losses on it each quarter; see who funds OpenAI.

Finance lease

A lease that works like buying the asset with borrowed money: it goes on the balance sheet, but the payments don’t count as capex. Some hyperscalers use finance leases heavily, so cash capex understates their spending.

Free cash flow

Operating cash flow minus capex: the cash a business generates after paying for its investment. How many of the largest AI buyers have negative free cash flow is one of the bubble indicators we track.

Gigawatt (GW)

A billion watts, now the unit in which AI data-center campuses are described. Nvidia’s guarantees cover leases for about 4.25 gigawatts of capacity for OpenAI in Ohio, and AMD’s agreements with OpenAI and Meta each cover up to 6 gigawatts of GPUs.

GPU (graphics processing unit)

A chip built to run many calculations at once, originally for graphics and now the standard AI accelerator. Nvidia’s data-center GPUs account for most AI computing; AMD is the main merchant alternative. See Nvidia’s customers and suppliers.

Gross margin

Revenue minus the direct cost of producing it, as a share of revenue. In the AI stack it shows who holds pricing power: Nvidia keeps most of each dollar, server makers keep little. See follow the money.

High-bandwidth memory (HBM)

Stacked DRAM placed next to an AI accelerator to feed it data fast enough; the reason memory became one of AI’s choke points. Read the full entry.

Hyperscaler

The handful of companies that build and run data centers at global scale, and whose capex drives the AI build-out. Read the full entry.

Inference and training

Training builds an AI model by processing huge amounts of data once; inference runs the trained model each time someone uses it. Inference grows with usage, which is why it is expected to become the larger share of AI computing over time.

Inventory provision

A charge a company takes when it expects to sell inventory for less than it cost, or not at all. Nvidia took a $4.5 billion charge for excess H20 inventory and purchase obligations after export controls cut demand.

Leases not yet commenced

Leases a company has signed for data centers that haven’t started yet, so they are not yet on its balance sheet or in its capex. Microsoft, Meta, Oracle, Alphabet and CoreWeave disclose more than a trillion dollars of them combined.

Liquid cooling

Cooling servers with liquid piped to the chips rather than with air. AI racks draw so much power that air cooling is no longer enough, which has made cooling suppliers such as Vertiv part of the AI stack; see the power layer.

Neocloud

An AI-focused cloud that rents GPU capacity, distinct from the traditional hyperscalers and usually financed with debt and long contracts. Read the full entry.

Nonvoting preferred stock

Shares that rank ahead of common stock and often carry investor protections but no votes. Amazon holds its Anthropic stake this way, which gives it economic exposure without control.

ODM and OEM

An original design manufacturer builds hardware to another company’s design (often in Taiwan); an original equipment manufacturer sells hardware under its own brand, such as Dell or HPE. Both are among Nvidia’s direct customers. See Nvidia’s customers and suppliers.

Operating cash flow

The cash a business generates from its operations before investment. Comparing capex with operating cash flow shows whether a company can pay for its build-out itself; the five largest AI buyers now spend nearly all of theirs. See is AI a bubble?.

Power purchase agreement (PPA)

A long-term contract to buy electricity from a specific generator, often for 20 years. Data-center buyers use them to secure power; generators use them to finance new or restarted plants. See the power layer.

Purchase commitments

Contracts to buy supplies or capacity in future, disclosed in the notes to the financial statements. Nvidia’s supply and capacity commitments have grown to hundreds of billions of dollars.

Related party

A company connected to the filer, for example through a large investment. Accounting rules require transactions with related parties to be disclosed, which is why Microsoft reports its revenue from OpenAI.

Remaining performance obligations (RPO)

Revenue a company is contractually owed but hasn’t yet earned: the closest thing in the filings to a measure of contracted future demand. Read the full entry.

Segment reporting

The breakdown of revenue and profit by business that companies must give when they manage the business that way. It is how the filings reveal AWS, Google Cloud, Microsoft Cloud and Nvidia’s data-center revenue.

Take rate

The share of each dollar a layer keeps as gross profit rather than passing on to its own suppliers. Our follow-the-money analysis measures it layer by layer.

TPU vs GPU

Google’s custom AI chip versus the general-purpose GPU, and why the hyperscalers’ custom silicon is the main long-term question for Nvidia. Read the full entry.

TTM (trailing twelve months)

The sum of a company’s latest four quarters. We use it for most figures because it smooths out seasonality and works across companies with different fiscal years.

Vendor financing

A seller helping its customer pay for what it buys; in AI it ranges from Nvidia’s lease guarantees to AMD’s customer warrants. Read the full entry.

Warrant (customer warrant)

The right to buy a company’s shares at a fixed price. AMD has issued warrants to OpenAI and Meta for up to 160 million shares each at a penny, vesting as they buy its GPUs: a discount paid in stock.

How to use these terms

Most of these words describe one of three things. Some describe the infrastructure itself: who owns the data centers (hyperscalers, neoclouds), what goes in them (accelerators, high-bandwidth memory, custom silicon) and who builds it (ODMs and OEMs). Some describe the accounting: how companies report what they have spent (capex), what it will cost them later (depreciation and useful life), what they are owed (remaining performance obligations) and what they have promised (purchase commitments, leases not yet commenced). And some describe the financing: how suppliers help customers pay (vendor financing, warrants, guarantees) and how long-term contracts lock in demand (power purchase agreements).

Reading a filing well means knowing which of those a number belongs to. A rise in capex is spending; a rise in RPO is demand that has been contracted; a rise in purchase commitments is a bet on demand to come. The AI stack map puts them together layer by layer, and the company pages show each one for a single company.

Every definition here is written for investors, not engineers: it explains what the term means for how a company earns money and what can go wrong. Where the technology matters to the economics, as it does for high-bandwidth memory, the full entry explains it; where it doesn’t, we leave it out.