Explained

Versus

Table-led comparisons of two terms — only after each Mechanism page exists.

On-premise vs cloud: empty data hall under construction with amber on the uninstalled power transformers that a dense AI rack needs

Explained

On-Premise vs Cloud for AI: Who Pays for the GPUs, and Where They Fit

On-premise means a company buys and runs its own servers; cloud means it rents them from a provider. For AI the choice is less about software than about who carries the cost of the GPUs and whether there is a room with enough power to plug them in.

Take-or-pay versus a purchase commitment: a stack of contracts beside a sealed server crate that has not been delivered

Explained

Take-or-Pay vs Purchase Commitment: What's the Difference?

A take-or-pay contract is a customer's promise to pay for reserved capacity whether or not every hour is used. A purchase commitment is a buyer's promise to pay a supplier later. Same words in a footnote can be opposite sides of the invoice.

Scale-up vs scale-out: photograph of an NVIDIA DGX GB200 rack with amber on the dense networking port cages where the scale-up domain meets scale-out

Explained

Scale-Up vs Scale-Out for AI: Who Gets Paid When Clusters Get Denser

Scale-up and scale-out are the two ways an AI cluster grows. Scale-up links GPUs inside one domain; scale-out links domains across Ethernet or InfiniBand. Denser racks grow the scale-up switch cheque; bigger halls grow the optics cheque.

InfiniBand vs Ethernet: macro photograph of a network switch with cables plugged in, in navy light, with two empty switch ports lit amber as the switch supply bind

Explained

InfiniBand vs Ethernet for AI: Who Gets Paid When the Fabric Changes

InfiniBand and Ethernet are the two network fabrics that link GPUs across racks in an AI cluster. Ethernet took about two-thirds of AI back-end switch sales in early 2026, but NVIDIA sells both. The fabric choice decides how the networking cheque is split; switch supply is the stuck step.

AWS vs Azure vs Google Cloud: photograph of a vast, mostly empty data centre hall in night-blue light, the bare floor lit amber as capacity not yet built

Explained

AWS vs Azure vs Google Cloud: Growth, Backlog and Capacity Compared

AWS is still the biggest cloud, but Azure and Google Cloud are growing faster. All three now hold more than $560 billion of signed backlog each and all three say the same thing: they cannot add capacity as fast as customers sign for it.

Explained

SoIC vs CoWoS: Height Versus Width — and Who Still Waits

SoIC stacks chips; CoWoS sits them beside HBM. TSMC sells both. The stuck step for a shippable AI GPU is still the CoWoS slot, not the 3D join.

Explained

Finance Lease vs Operating Lease: Same Data Centre, Two Ledgers

A finance lease is booked like buying with borrowed money; an operating lease is booked as rent. Same building, same rent, different lines. Microsoft's new 25-year building life moves leases off its capex line.

Reserved vs on-demand GPU: a dark aisle of identical locked rack cabinets, one cabinet door lit amber as the reserved rack that bills whether or not it runs

Explained

Reserved vs On-Demand GPUs: Who Pays for the Idle Hour?

A reserved GPU bills for its whole term; an on-demand GPU bills only while it runs. Neoclouds live on the first, the second sets the price signal, and in 2026 the discount for committing has been shrinking.

Capex vs depreciation: stacks of still-bagged compute boards in front of supercomputer racks, one stack lit amber, with the words bought now, expensed later

Explained

Capex vs Depreciation: Why AI Spending Hits Profit Later

Capex is the cash out today; depreciation is the bill that reaches profit over the next five to six years. The five biggest AI buyers spent about four dollars in cash for every dollar of depreciation they booked in their latest quarters, and the clock only starts when a powered hall switches on.

Explained

AI Tokens vs GPU Hours: Same Chip, Different Bills

Tokens are how a cloud bills the model. GPU hours are how it bills the card. They only match at a conversion ratio — billed tokens per live GPU-hour.

Photoreal CoWoS full silicon interposer glowing amber beside a dark EMIB organic substrate — the full sheet is the bind

Explained

CoWoS vs EMIB: Two Ways to Join an AI GPU — and Who Gets Paid

CoWoS puts an AI GPU and HBM on a full interposer. EMIB embeds small silicon bridges in the substrate. The stuck step is still the CoWoS slot, not the second recipe.

Hyperscale vs colocation data center: caged colocation suites with yellow fibre on the left, a vast uniform hyperscale hall on the right, with the words $280 vs $157 per kW, colocation vs hyp

Explained

Hyperscale vs Colocation Data Centers: Same Power, Different Rent

A hyperscale data center is a huge hall for one cloud or AI giant; a colocation data center rents cages, racks and connections to many tenants. Per kilowatt, colocation pays far more. Per lease, hyperscale lasts twice as long, and power scarcity is now repricing the big leases.

What Versus pages do

A Versus Explained page compares two mechanisms side by side. It is table-led, shorter (about 1,000–1,400 words), and never ships before both Mechanism pages exist.

URL shape: /explained/<a>-vs-<b>/. Use it when the money question is which path wins, not what one part is.

High-demand comparisons

Neocloud vs hyperscaler (and neocloud vs hyperscale)
HBM vs GDDR / HBM vs DDR
EMIB vs CoWoS

More side-by-sides: EUV vs DUV · TPU vs GPU · HBM vs GDDR7 · HBM vs DDR5.

Start from Mechanism if you need the definition first.