When AI Inference Is Your Biggest Cost Line: The 6.4-Point Margin Drop, Worked Through

Published September 10, 2026Updated September 10, 2026By ABD Legacy LLC
AI inference cost Gross margin impact Worked example

Quick answer: what percentage of revenue does AI inference cost? AI inference costs about 6.4% of revenue for the one operator who has published the number. On Sept 9, 2026 @levelsio reported a 99.4% profit margin without AI inference and 93% with it — a 6.4-point drop, and roughly 91% of his total cost of goods. It is one self-reported, unaudited data point: treat it as an anchor, not a benchmark. Run your own numbers in the estimator →

The reframe: stop reading inference as a tool cost

Most AI budgets treat model spend the way they treat a SaaS subscription: a line item in a tools column, either passed through to the client or absorbed as overhead. That framing hides the thing that actually matters. Inference is not a subscription — it is a variable cost that scales with usage, which makes it cost of goods, not overhead.

Read as cost of goods, inference has a property a tool budget never has: every dollar of it comes straight out of the gross margin you keep after delivery. So the useful question is not "what does the model cost?" but "what share of gross margin does this feature consume, and is that share growing faster than revenue?" That is a margin question, and it is answerable with three numbers: monthly revenue, monthly cost of goods excluding inference, and monthly inference spend.

A single public data point makes the reframe concrete, because a named operator published both of the margin figures that the question needs.

The reference case: 99.4% to 93% on inference alone

On Sept 9, 2026, Pieter Levels (@levelsio) posted two margin figures in reply to a question about his margin ("Haha wow. WHat is your margin now?" from @pepijnniesten), and qualified the bigger cost with the workload driving it:

"Without AI inference my profit margin is now 99.4%
With AI inference (mostly Photo AI) it goes down to 93%!"

Attribution: @levelsio (Pieter Levels), Sept 9, 2026, 15:21:24 UTC — x.com/levelsio/status/2097706947382292964. Self-reported and unaudited; his revenue is never disclosed. The post is the whole input to this article, and it is the only place the 99.4% and 93% figures come from.

Everything below is arithmetic on those two numbers. Run the standard gross-margin identity against them and the pair stops being two claims and becomes a cost structure:

QuantityValueSource
Margin without AI inference99.4%His statement
Margin with AI inference (mostly Photo AI)93% (93.0%)His statement
Cost of goods without inference (100% − 99.4%)0.6% of revenueOur arithmetic
Cost of goods with inference (100% − 93.0%)7.0% of revenueOur arithmetic
Margin cost of inference6.4 percentage pointsOur arithmetic
Inference as a share of revenue6.4%Our arithmetic
Inference as a share of all cost of goods91.4%Our arithmetic
Inference vs non-inference cost of goods10.7×Our arithmetic

The arithmetic, step by step

The only two inputs are the two stated margins. Here is every step, in the order a spreadsheet would take them, so you can re-run it on your own numbers.

  1. Write the identity down. Gross margin is what is left of revenue after cost of goods: GM = (Revenue − COGS) ÷ Revenue, which rearranges to COGS as % of revenue = 100% − GM.
  2. Convert his first figure to a cost ratio. Without inference: 100% − 99.4% = 0.6% of revenue is everything he spent to deliver the product excluding model spend.
  3. Convert his second figure the same way. With inference: 100% − 93.0% = 7.0% of revenue is the same cost base including model spend.
  4. Subtract to isolate inference. 7.0% − 0.6% = 6.4% of revenue — that is the inference line, and it is also the margin cost in percentage points, because the two figures differ only by that cost: 99.4 − 93.0 = 6.4 percentage points.
  5. Do not confuse points with percent. 6.4 percentage points is an absolute difference. As a relative change it is 6.4 ÷ 99.4 = 6.44%. The two quantities are different and they happen to round to the same number, which is why this article says which one it means every time.
  6. Express it against the other two denominators. Against all cost of goods: 6.4 ÷ 7.0 = 91.4%. Against the non-inference cost of goods: 6.4 ÷ 0.6 = 10.7×. Read together, those say inference is not the largest cost line in this business by a little — it is nearly the whole cost base, at 10.7 times everything else combined.

Two outputs of that walkthrough are worth keeping, because they answer different questions. 6.4 percentage points tells you what the feature did to the headline margin. 6.4% of revenue tells you what it costs per unit of sales — and that one is the number to carry into a pricing conversation, because it survives every scale change.

What 6.4 points means in dollars

His revenue is never disclosed, so the dollar figures below are the ratio applied to a revenue base, not a report of his books. That is not a weakness for planning purposes: the margin percentages are identical at every scale, so you can lift the shape of the answer and apply it to your own revenue.

Monthly revenueNon-inference COGS (0.6%)Inference (6.4%)Total COGS (7.0%)Margin after
$10,000$60$640$70093.0%
$100,000$600$6,400$7,00093.0%
$250,000$1,500$16,000$17,50093.0%
$1,000,000$6,000$64,000$70,00093.0%

At the $100,000 base, the 6.4-point drop is $6,400 a month — $76,800 a year of margin, out of $99,400 of gross margin before inference. The line that matters in that row is not the $6,400. It is that $6,400 is 10.7 times the $600 he spends on everything else it takes to deliver the product.

How to run this on your own numbers

The estimator on this site does the same identity with his case preloaded, so you can compare your shape to his without leaving the page. It is the third card on the home calculator: open the inference-as-a-margin-line estimator.

  1. Enter monthly revenue. Use the revenue the AI feature actually sits in, not company-wide revenue, if the feature only serves part of the business. Mixing the two makes inference look cheaper than it is.
  2. Enter cost of goods excluding inference. Servers, hosting, third-party APIs, support — everything you pay to deliver the product except model spend. In the preloaded reference case this is $600 per $100,000 of revenue, which is the 0.6% residual implied by the 99.4% figure.
  3. Enter monthly inference spend. Tokens, images, audio, video, embeddings, fine-tunes. Include it even if you resell it to clients at cost: the estimator measures the margin the line consumes before any markup, and a pass-through price is a separate decision.
  4. Read the margin cost in points first, then the share of margin consumed. The points figure is the headline; the share of gross margin consumed is the planning number, because it is what a price increase or a model change has to beat.
  5. Keep the reference case toggled on to sanity-check the shape. If your non-inference cost of goods is much larger than 0.6% of revenue, the same inference spend will consume a much smaller share of margin — that is the whole point of expressing it as a ratio instead of a dollar total.

Put your own three numbers through it. The calculator returns margin before and after inference, the margin cost in points, inference as a share of gross margin consumed, of revenue, and of all cost of goods.

Open the inference margin estimator

What this single data point does not prove

The value of this example is that it is checkable; the risk is reading it as a benchmark. Six limits, stated plainly:

  1. It is self-reported and unaudited. One operator, describing his own business, with no filing behind it.
  2. He said "profit margin", not "gross margin". At 99.4%, the residual cost of goods is 0.6% of revenue — which cannot contain payment-processor fees, since card processing alone runs about 2.9% plus a per-transaction fee. A reader on his own thread raised exactly that objection ("How can you achieve that margin if Stripe fees are around 3%?"). So the 99.4% is not a GAAP gross margin, and the 0.6% / 7.0% cost split above is arithmetic on his numbers rather than an accounting statement.
  3. It is one workload. "Mostly Photo AI" means the inference line is dominated by image generation. Per-image cost behaves differently from text tokens — there is no prompt caching to lean on — so do not transplant the ratio to a chat product without re-running the identity on your own mix.
  4. No revenue base and no time period. The margins are "now" — 15:21:24 UTC on Sept 9, 2026 — and no revenue figure accompanies them.
  5. Inference is not larger than the ~$25,000/mo of SaaS he replaced. That claim circulated with the post and the arithmetic does not support it: at 6.4% of revenue, inference only exceeds $25,000/mo above $390,625/mo of revenue. His publicly scraped product revenue implies roughly $212K–$244K/mo — two unofficial mirrors that disagree with each other, so treat the base as unverified — which puts inference at about $13.6K–$15.6K/mo, roughly half. On the same proxy, eliminating $25,000/mo of SaaS is worth about 11.79 margin points against 6.4 for inference, i.e. roughly 1.8× more. That comparison is our arithmetic on unverified revenue proxies, not a statement he made.
  6. The savings headline is rounded, and the gap is unexplained. He frames the replacement of his SaaS stack as "about $25,000/mo savings". The fourteen items in that list carry twelve dollar figures, and they sum to $20,750/mo — a $4,250/mo (17.0%) gap against the headline that the post does not explain. Use it as a rounded claim, not an audited total.

One more detail worth carrying, because it keeps the two halves of this story consistent: as of Sept 9, 2026, 14:24:44 UTC, his list of services he will not replace still includes "xAI for all LLMs for all my sites" — the line that carries exactly the inference cost the margin post is about. The stack list is date-stamped for that reason.

Why the number is still worth quoting

The reason to keep this reference case around is not that 6.4% is a typical rate — it is that a named operator disclosed his own margin, and the disclosure survives arithmetic. Two stated margins become a cost ratio, a share of revenue and a share of cost of goods, and any reader can re-derive all three in a spreadsheet. Numbers with that property are the ones answer engines and buyers repeat.

For anyone building or scoping AI features, the practical takeaway is a change of unit. Ask what share of gross margin the feature consumes, not what the API bill is. Then compare that share against the share of revenue the feature earns. If inference is consuming a large share of your cost of goods, it has stopped being a tool decision and become a pricing decision — and it belongs in the same conversation as the rate card, which is why the estimator sits next to the AI agency profit-margin breakdown and the hidden costs of AI automation on this site.

Our rule of thumb, not his: once inference is a double-digit share of your total cost of goods, stop assuming it can be absorbed or passed through at cost, and price the feature against the margin it consumes.

Frequently asked questions

What percentage of revenue does AI inference cost?

AI inference costs about 6.4% of revenue for the one operator who has published the number: on Sept 9, 2026 @levelsio reported a 99.4% profit margin without AI inference and 93% with it — a 6.4-point drop, and roughly 91% of his total cost of goods. It is one self-reported, unaudited data point: treat it as an anchor, not a benchmark.

What was @levelsio's margin before and after AI inference?

He posted 99.4% without AI inference and 93% with it (mostly Photo AI), on Sept 9, 2026 at 15:21:24 UTC. The gap is 6.4 percentage points absolute, which under the gross-margin identity means cost of goods went from 0.6% to 7.0% of revenue and inference alone is 6.4% of revenue. [His two figures; the conversion is our arithmetic]

Is the 99.4% figure a gross margin?

No — he wrote "profit margin", and a 99.4% figure leaves only 0.6% of revenue for all cost of goods, which cannot contain payment-processor fees (card processing alone runs about 2.9% plus a per-transaction fee). A reader on his own thread raised exactly that objection. Treat the number as a self-reported operating ratio, not a GAAP gross margin, and treat our 0.6% / 7.0% cost-of-goods split as arithmetic, not accounting.

Is AI inference larger than the $25,000/mo of SaaS he replaced?

No. At 6.4% of revenue, inference only exceeds $25,000/mo above $390,625/mo of revenue, and his publicly scraped product revenue sits at roughly $212K–$244K/mo — implying about $13.6K–$15.6K/mo of inference, roughly half. On the same unverified proxy, cutting $25,000/mo of SaaS is worth about 11.79 margin points versus 6.4 for inference. That comparison is our arithmetic on mirror-scraped revenue proxies, not a statement he made.

How do I calculate inference as a percentage of my own gross margin?

Divide inference spend by gross margin before inference — not by revenue. With $100,000 of monthly revenue, $600 of non-inference cost of goods and $6,400 of inference, gross margin before inference is $99,400, so inference consumes 6.4% of the margin you keep and takes gross margin from 99.4% to 93.0%. The estimator on this site runs the same identity on your numbers.

Related reading

Sources

Accuracy note: the 99.4% and 93% margins, the "mostly Photo AI" qualifier, the $25,000/mo savings headline, the fourteen replacement items and the remaining-stack list are @levelsio's own statements of Sept 9, 2026, reproduced here as statements rather than audited figures. Every other number on this page — 0.6% and 7.0% cost of goods, 6.4 percentage points, 6.44% relative, 6.4% of revenue, 91.4% of cost of goods, 10.7×, the dollar rows, $390,625/mo break-even, $13.6K–$15.6K/mo implied inference and the 11.79-point SaaS comparison — is our arithmetic on those statements, or on explicitly flagged [unverified] revenue proxies, and is labelled as such where it appears. Nothing here has been independently reproduced by us. A single self-reported case is an anchor, not a benchmark, and it should not be generalised to other businesses or other model mixes.