Blog
Analysis written from the same sourced, dated figures the directory publishes. Every number in a post is read live, so a post cannot go stale without its page saying so.
Why the same GPU has different prices
The same GPU costs different amounts at different providers because the rate buys more than the card: the host, the network, the terms and the region.
- GUIDE · LLM costs
What a token is and why it sets your bill
A token is a chunk of text, about four characters of English, and an LLM API bills every token it reads and every token it writes at separate rates.
27 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · Renting GPUs
Rent for a week or commit for a year
A reserved GPU pays off when the hours you will really use, priced at the on-demand rate, cost more than the commitment. Your utilisation sets the line.
27 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · LLM costs
Input and output token prices, explained
Input and output tokens carry separate per-million rates, output is commonly dearer, and a blended rate only means something for one fixed mix of the two.
27 OCT 2026 · 5 MIN · 4 SOURCES - NOTE · Renting GPUs
Idle GPU time is the expensive part
Idle GPU hours, overnight, between jobs or waiting on data, bill at the full on-demand rate. Find and stop them before shopping for a lower hourly rate.
27 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · LLM costs
How to estimate a monthly LLM bill
Count requests, measure average input and output tokens per request, multiply each by its rate per million, and add the hidden calls before trusting the total.
27 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · Renting GPUs
Compare GPU providers fairly: same card, same terms, same unit
A fair GPU provider comparison holds three things fixed: the exact card, the commitment and the unit. Then it adds what each rate leaves out.
27 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · LLM costs
Batch pricing: when it fits and when it does not
A batch tier prices tokens lower in exchange for a delayed answer, so it fits work nobody is waiting on and fails anything a person or a pipeline is blocked by.
27 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · Renting GPUs
What a month of GPU time costs
A month of GPU time is the hourly rate times 730 hours if the machine never stops. Most bills are smaller or larger than that for reasons you can count.
26 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Choosing a GPU
SXM or PCIe for your workload
SXM boards link GPUs directly and suit models split across cards; PCIe cards sit in an ordinary slot and suit single-card work at a separate price.
26 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Renting GPUs
How to read a GPU price page before you rent
Read a GPU price page in five passes: the unit, the commitment, the exact card, the region and what the rate includes. Only then read the hourly figure.
26 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Renting GPUs
On-demand, reserved or interruptible: which GPU rental fits
On-demand suits work you cannot predict, reserved suits load you can, and interruptible suits jobs that save their state and restart without a person.
26 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Choosing a GPU
A GPU for fine-tuning or for inference: different needs
Full fine-tuning needs memory for gradients and optimiser state on top of the weights; inference needs it for the cache. Size the GPU for the job you run.
26 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Renting GPUs
Egress and storage: the extras on a GPU bill
The two extras most likely to surprise a GPU renter are storage that bills while the instance is off and fees for moving data out. Ask about both first.
26 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Renting GPUs
A checklist before your first GPU rental
Before a first GPU rental, size the memory, pick the card, price a month, read the terms, ask about storage and egress, and set a way to stop paying.
26 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Choosing a GPU
24 GB, 48 GB or 80 GB: which memory tier
Pick the smallest memory tier that holds your weights plus measured headroom: 24 GB for small models, 48 GB for mid-size, 80 GB for 70B in 8-bit.
26 OCT 2026 · 5 MIN · 4 SOURCES - PRICE ANALYSIS · GPU market data
RTX 4090 rental prices across providers
RTX 4090 24GB on-demand rates from marketplaces and managed providers, how one rate per seller is chosen, and what a 4090 seller should check.
25 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Choosing a GPU
Using an RTX 4090 for LLM work: where it stops
An RTX 4090 has 24 GB: room for a 7B model in 16-bit or a 32B model at 4-bit. Past that it stops, and a rental card with more memory is the next step.
25 OCT 2026 · 5 MIN · 4 SOURCES - PRICE ANALYSIS · GPU market data
How many providers rent each GPU
The count of providers with a current on-demand rate for each GPU, card by card, and what a short list of sellers means for a provider entering it.
25 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · GPU market data
H100 SXM rental prices across providers
Where H100 SXM 80GB on-demand rates sit across providers, read from each seller's page, and what a seller below or above the range should check.
25 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Choosing a GPU
How much GPU memory a model needs for its weights
Weights need the parameter count times the bytes per parameter: a 70B model is about 140 GB in 16-bit, 70 GB in 8-bit and 35 GB in 4-bit, before any cache.
25 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · GPU market data
Consumer cards in the rental market: 4090, 5090 and 3090
Rental rates for the RTX 5090, 4090 and 3090 side by side, what 32 GB and 24 GB hold, and why a seller of consumer cards should name each one exactly.
25 OCT 2026 · 4 MIN · 4 SOURCES - PRICE ANALYSIS · GPU market data
Blackwell rental prices so far: B200, B300 and GB200
What sellers list for Blackwell rental so far: B200, B300 and GB200 rates, how many providers list each, and why a thin market reads differently.
25 OCT 2026 · 5 MIN · 4 SOURCES - PRICE ANALYSIS · GPU market data
A100 pricing: SXM, PCIe and the unlabelled rate
A100 rates sit in a row for the SXM 80GB and a row for pages naming no form factor. What each holds, and why an unlabelled rate is hard to compare.
25 OCT 2026 · 4 MIN · 3 SOURCES - METHOD · Working with infratoapp
Why paying does not change where you rank
The query that orders every public list does not join the billing table, and the audit is published. Paid placement is labelled and sits outside the list.
24 OCT 2026 · 5 MIN · 4 SOURCES - NOTE · Working with infratoapp
Seats and roles: who on your team can see what
The person who claims becomes admin, colleagues at the domain join read-only, and only organisation admins see plans and invoices. Editors and analysts do not.
24 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Working with infratoapp
Reading your thirty-day totals: views, comparisons, clicks
Every claimed listing shows thirty-day totals of views, comparisons, outbound clicks and opt-ins against the thirty days before. Counts under 5 are withheld.
24 OCT 2026 · 5 MIN · 3 SOURCES - NOTE · Working with infratoapp
Pro Profile: what it is, and what it does not buy
Pro Profile is a reviewed block in your own words at the top of one product page, labelled as yours. It changes no fact and no rank, and it is sold per product.
24 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Working with infratoapp
Intent signals: what you can see, and why it is counts
Intent signals show counts of what readers do on your product page, withhold any count under 5, and name a reader only when that reader chose to share.
24 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Working with infratoapp
Featured placement: one card, one layer, outside the ranking
Featured is one labelled card on one layer page, outside every ranked list, only on a layer you are listed in. Unsold, the slot is gone. It cannot move a rank.
24 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Working with infratoapp
Disputing a value we hold: what evidence works
A dispute works when it points at a page we can re-read that says something different. An assertion without a page leaves the published value where it is.
24 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Working with infratoapp
How buying works here: an invoice from a person
There is no checkout and no card on file. You email, a person creates an order and sends an invoice, and the plan switches on when the invoice is marked paid.
24 OCT 2026 · 5 MIN · 3 SOURCES - NOTE · Vendor marketing
When your docs and your pricing page disagree
A rate on the pricing page that differs from the one in the docs gives a buyer no way to tell which is right. Keep one source for prices and build from it.
23 OCT 2026 · 5 MIN · 4 SOURCES - METHOD · Working with infratoapp
What infratoapp reads from your pricing page, and how often
A crawler reads your own pricing page on a schedule, keeps the sentence each rate came from with its URL and date, and hides any price older than 21 days.
23 OCT 2026 · 5 MIN · 4 SOURCES - METHOD · Vendor marketing
Structured data for a GPU pricing page
Mark up each rate as a Product with an Offer that states price, currency, unit and a date, and never claim more in markup than the visible page shows.
23 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Vendor marketing
Date your prices so every quote carries its age
A price with no date could be from any month. Show when your rates last changed, in text and in markup, and change that date only when a price moves.
23 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Vendor marketing
Writing a comparison page that does not read as an ad
A comparison a buyer believes uses the same attributes for every row, says where you lose, discloses that you wrote it, and dates every figure.
23 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Working with infratoapp
Claim your listing on infratoapp: what happens, in order
You prove you control your company domain with a work email, you get a workspace as its admin, and you can then edit some fields and request changes to others.
23 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Vendor marketing
Case studies a technical buyer will believe
A case study an engineer believes names the workload, the scale, the before and after, what did not work, and who the customer is, with a date.
23 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Model makers
A model page an assistant can quote correctly
Assistants quote pages that state facts plainly: an answer first, exact names, dated figures and structured data. None of that guarantees a citation.
23 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · Model makers
Weights size and the GPU it fits: publish the arithmetic
Weights need about the parameter count times the bytes per parameter. A model page that works this out for each precision tells a buyer which GPU to rent.
22 OCT 2026 · 5 MIN · 4 SOURCES - NOTE · Model makers
Say when a model retires and what replaces it
A retirement notice needs two facts: the date the model stops answering and the model that replaces it. Both belong on the model page, not only in an email.
22 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Model makers
Release notes that say what changed and what to retest
A model update can change behaviour, the prompt format, the tokenizer or the price. Release notes should list each change and the test a buyer must run again.
22 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Model makers
Open weights: list where the model can be served
Published weights mean many sellers. The page buyers want is a dated list of hosts, the build each one serves, and the limits each one enforces on the model.
22 OCT 2026 · 5 MIN · 4 SOURCES - NOTE · Model makers
Name your models so every seller spells them alike
When sellers spell one model several ways, price tables split it into separate rows. Publish one canonical name and a grammar for its variants, and keep both.
22 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Choosing a GPU
The L40S for inference: when 48 GB is enough
The L40S has 48 GB of graphics memory, enough for a 7B to 13B model in 16-bit or a 70B model squeezed to 4-bit, with the cache as the limit to test.
22 OCT 2026 · 4 MIN · 4 SOURCES - GUIDE · The AI stack
Guardrails and the security layer
A guardrail checks what goes into a model, what comes out, and what an agent is about to do. Each check catches different failures and costs latency to run.
22 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Choosing a GPU
The AMD MI300X as an alternative to NVIDIA
The MI300X has 192 GB against the H100's 80 GB. The question to settle before price is whether your software stack runs on AMD as it runs on NVIDIA.
22 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Choosing a GPU
When a B200 is worth asking about
A B200 is worth asking about when your model needs more than 141 GB on one card or more than one H200. Ask about stock, unit and terms before price.
21 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Vendor marketing
What an MLPerf result does and does not tell a buyer
An MLPerf result is what one whole system scored on one benchmark and scenario. It says nothing about price, availability or the system you would rent.
21 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · The AI stack
Observability and evals for LLM apps
Tracing records what each model call did; evaluation scores whether it was right. An LLM app needs both, and they answer different questions at different times.
21 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · GPU market data
MI300X rental prices and who sells them
Who lists the AMD MI300X 192GB on demand, the lowest and highest rates they publish, and what a seller of an AMD card is compared against.
21 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Renting GPUs
Marketplace or managed GPU cloud
A marketplace rents you hosts listed by many owners at prices the market sets; a managed cloud runs its own fleet. Pick by what you need when a machine fails.
21 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · GPU market data
L40S rental prices across providers
L40S 48GB on-demand rates across sellers, set beside the L40 that shares its memory size, and the naming checks an L40S seller should run first.
21 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · Vendor marketing
Benchmark claims: publish the method with the result
A benchmark result is checkable only with its method beside it: the hardware, software versions, model, precision, batch size, sequence lengths and runs.
21 OCT 2026 · 5 MIN · 4 SOURCES - PRICE ANALYSIS · GPU market data
B200 rental prices across providers
B200 192GB on-demand rates as listed by providers, read against Hopper prices, with the checks a Blackwell seller should run on its own page.
21 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference and serving
Serverless or dedicated inference
Serverless bills per call and suits uneven traffic; dedicated bills per hour and suits steady load. The deciding figure is how busy the GPU would be.
20 OCT 2026 · 5 MIN · 5 SOURCES - PRICE ANALYSIS · GPU market data
H200 rental prices across providers
The lowest and highest listed H200 141GB on-demand rates, how they compare with the H100, and the memory case a seller can make for the card.
20 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Choosing a GPU
H100 or H200 for inference: what the memory difference buys
The H200 has 141 GB against the H100's 80 GB. For inference that extra memory is what you pay for, and it pays only when your model or cache needs it.
20 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Inference and serving
A gateway or router for model calls
A gateway sits between your code and model providers to handle keys, limits, logging and fallback; a router picks which model gets each call. Both add a hop.
20 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · LLM costs
Cached input tokens: when they lower the bill
A cached input rate bills a repeated prompt prefix for less, but only where the seller publishes one; everywhere else the same tokens pay the full input rate.
20 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · Inference and serving
API or self-host: where the break-even sits
Self-hosting beats an API when your monthly token bill exceeds the rental month of the GPUs needed to serve the same traffic, plus the cost of running them.
20 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · The AI stack
Agent frameworks: what to compare
Compare agent frameworks on how explicit the control flow is, how state persists, how tools and MCP connect, and what you can see when a run goes wrong.
20 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Choosing a GPU
A100 or H100: which to rent
Rent the A100 when the model fits and the job is not bound by 8-bit maths; rent the H100 when you need Hopper's FP8. Price both from the listed rates.
20 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · MCP and agent skills
Test an MCP server without risking your data
Test a new MCP server against throwaway data with a limited key in a separate account or folder, and watch every tool call before trusting it.
19 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Vendor marketing
Structure a pricing page for search and assistants
Put every rate in a plain HTML table, quote one unit, give sections stable anchors, date it, and keep no price behind a tab, a toggle or a calculator.
19 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Inference and serving
Quantization: what you trade for the smaller model
Quantization stores weights in fewer bits so a model fits on fewer or smaller GPUs. You give up some precision, and only your own tests say what that costs.
19 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · MCP and agent skills
Publish an MCP server people can find
A findable MCP server names its credentials in the README config block and its service in its description, carries a licence and shows recent pushes.
19 OCT 2026 · 5 MIN · 4 SOURCES - NOTE · MCP and agent skills
Official or community MCP server: how to tell
An official MCP server is published by the company whose product it reaches; check the publishing account and the vendor docs, and read the code anyway.
19 OCT 2026 · 5 MIN · 3 SOURCES - NOTE · Vendor marketing
What llms.txt does for an infrastructure vendor
llms.txt is a plain-text index of your key pages. Google has said its search does not use it, some tools may read it, and it takes about an hour to write.
19 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference and serving
The KV cache: why context costs memory
Every token in a request leaves keys and values in GPU memory until the reply ends, so long context and many users fill a card long before the weights do.
19 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Vendor marketing
How assistants pick the GPU prices they quote
Nobody outside the labs knows how assistants weigh sources. A price is easier to quote when it is readable as text, tied to its page and dated.
19 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · MCP and agent skills
Claude skills: what they are and where to find them
A Claude skill is a folder with a SKILL.md file: a name, a description and instructions a model loads. Here is what that means and where skills come from.
18 OCT 2026 · 4 MIN · 3 SOURCES - GUIDE · MCP and agent skills
What an MCP server is, in plain terms
An MCP server is a small program that lets an AI assistant use tools and data outside the chat, such as files, a database or a ticket tracker.
18 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · MCP and agent skills
MCP servers by what they reach
Sort MCP servers by the system they touch, a database, a code host, files or a chat tool, because each kind needs a different credential and caution.
18 OCT 2026 · 5 MIN · 3 SOURCES - NOTE · MCP and agent skills
MCP server credentials and what can go wrong
An MCP server acts with the credential you give it, so the scope of that key, not the server description, sets the limit on what a model can do.
18 OCT 2026 · 5 MIN · 3 SOURCES - METHOD · MCP and agent skills
How the MCP directory is built and what it leaves out
Every MCP server listed is read from its own repository, credentials come from its README config, and the count of what was held back is published.
18 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · MCP and agent skills
Evaluate an MCP server before you install it
Before installing an MCP server, read its tool list, its configuration block and its recent history, and decide what credential it actually needs.
18 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · MCP and agent skills
Choosing agent skills from a directory
Pick an agent skill by reading its SKILL.md, checking who publishes it and where it lives, and testing it on one task you already know how to do.
18 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · MCP and agent skills
Agent skills or MCP servers: which do you need
Use an MCP server when the model must reach a system it cannot see, and a skill when it already has the access but needs to know how to do the task.
18 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
What buyers compare before they choose an LLM API
An LLM API is checked on the token rate, then context length, then rate limits, then what happens when the model retires. Skip one and the buyer can leave.
17 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · GPU cloud vendors
What buyers compare before they rent a GPU
Buyers start with the hourly rate and finish with availability, network and terms, and that second list is where a provider keeps or loses the sale.
17 OCT 2026 · 4 MIN · 4 SOURCES - GUIDE · GPU cloud vendors
SXM or PCIe: name the card exactly on your price page
An H100 SXM and an H100 PCIe share a name and are sold at different prices. A price page that says only "H100" leaves the buyer to guess which one is on offer.
17 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Model makers
State your model license in plain words
A license name tells a buyer where to look, not what they may do. Put the permissions in a short can and cannot list beside the name, dated, on the model page.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
State the context window on every model page
A filter on context window cannot include a model with no stated window. With no figure on the page, the model drops out of the list.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
Resellers: say whose model you are selling
A seller of someone else's model should print the maker and exact version, so one model is not listed as two and buyers can compare you.
17 OCT 2026 · 5 MIN · 4 SOURCES - METHOD · GPU cloud vendors
Regional GPU pricing without confusing the comparison
A GPU priced differently by region needs a headline rate and a labelled regional table, or buyers compare one provider’s cheapest region with another’s dearest.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
Rate limits belong on the pricing page, not only in the docs
A token price without the rate limit attached is half a price. Buyers hit limits at launch, so state them by tier where the rates are, in the same units.
17 OCT 2026 · 5 MIN · 3 SOURCES - NOTE · Model makers
Put the parameter count on the model card
Every estimate of the memory a model needs starts with its parameter count. Print it on the model card as a number, with the precision the weights ship in.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
Publish input and output token prices side by side
A token price built from the input rate alone is a different number from a bill. Put both rates next to each other, per million tokens, for every model.
17 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · GPU cloud vendors
Per card or per node: pick one unit for your GPU prices
A node price beside a per-card price cannot be compared until someone does a division, and many buyers will not. Publish the rate per GPU per hour.
17 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · Inference vendors
One model, many sellers: how far token prices spread
Some models we price are sold by more than one seller. For those, buyers compare sellers, not models, and a page has to say what the dearer rate buys.
17 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · GPU cloud vendors
Label every GPU rate: on-demand, reserved or interruptible
On-demand, reserved and interruptible are three products. Put the label beside every rate, or buyers will compare your cheapest row with a rival’s dearest.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
A checklist for an inference pricing page
What an inference pricing page needs so buyers can compare it fairly: exact names, both rates, cached and batch terms, context, limits and retirement dates.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Model makers
How buyers find a new model, and where yours is missed
Buyers meet a new model through search, assistants, directories and price tables. A model with no price, context figure or license drops out at the filter.
17 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · GPU cloud vendors
A checklist for a GPU pricing page buyers can compare
A GPU pricing page buyers can compare names the card exactly, quotes per GPU per hour, labels the commitment and states egress, regions and access mode.
17 OCT 2026 · 4 MIN · 4 SOURCES - NOTE · GPU cloud vendors
Say what is in stock: GPU availability on a pricing page
A rate for a GPU nobody can rent is a rate nobody can pay. An availability line beside the price tells buyers whether the card exists for them today.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · GPU cloud vendors
Your GPU listing is wrong: how to correct it
A wrong value on a GPU listing is fixed one of two ways: edit the fields a claimed vendor controls, or dispute the value with evidence. Here is which is which.
17 OCT 2026 · 5 MIN · 4 SOURCES - GUIDE · Inference vendors
Explaining cached input pricing so buyers can use it
A cached input rate is worth something only if buyers know what counts as cached, how long it lasts and what they must do. A bare rate says none of it.
17 OCT 2026 · 5 MIN · 3 SOURCES - NOTE · GPU cloud vendors
Egress fees: the GPU cost buyers ask about before they sign
Buyers who have paid to move data out of a cloud ask about egress first. A GPU page that is silent on it is read as hiding a charge, whether or not one exists.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · GPU cloud vendors
How to describe interruptible GPU instances honestly
Before choosing an interruptible GPU a buyer needs the reclaim warning, what happens to the job and how the rate is set, not just the size of the discount.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
Deprecation notices buyers can plan around
A model retirement notice is useful when it gives a date, a named replacement and a migration note in one place. Buyers read the policy before they commit.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Model makers
Context length and modalities: the two facts buyers filter on
Buyers sort models by how much text they accept and what kinds of input they take. State both in tokens and plain lists, or the model falls out of the filter.
17 OCT 2026 · 5 MIN · 3 SOURCES - GUIDE · Inference vendors
Batch tier pricing: state the trade-off
A batch tier trades speed for price. Buyers sort their calls into those that can wait and those that cannot, so state the window, the limits and the rate.
17 OCT 2026 · 5 MIN · 3 SOURCES - PRICE ANALYSIS · GPU pricing
The same GPU, at very different prices
A GPU is a commodity. Renting one is not priced like a commodity — on the cards enough providers rent to compare, the dearest typically charges 3.7× the cheapest, and on the widest gap far more than that.
16 SEPT 2026 · 4 MIN · 4 SOURCES
The digest
Leave an email and we will use it the day the digest starts sending. It is not sending yet.
NOTHING SENDS YET — WE ARE ONLY SAVING ADDRESSES