0
Skip to Content
Ferrous Pro
New Page
Ferrous Pro
New Page
New Page
Ferrous
The Gap What We Do For Investors Why Us Get your free benchmark
Compute Economics for Venture-Backed AI Companies

We negotiate your model API commits and GPU renewals.
Then we make them fundable.

Cloud negotiation firms won't take your call below enterprise scale — and none of them touch your OpenAI, Anthropic, or Bedrock commits at all. Your CFO firm models the margin problem but never sits across the table. Ferrous negotiates the AI contracts your margin story depends on, then builds the unit-economics narrative your raise requires. Buyer-side only. Seed to Series C.

Get your free 48-hour benchmark ↗ See what you should pay

Send one contract or invoice. Get a two-page benchmark memo in 48 hours. Free. No call required.

52% vs 60%+
average gross margin for AI product builders — against the 60%+ bar investors now cite at Series A. Your compute contracts decide which side you're on.
ICONIQ State of AI Jan 2026 · Carta Q1 2026 · CRV
4–5×
what AI contract renewals are reportedly coming back at — the vendor knows every price they've ever charged; you know one
Reported enterprise renewals · 2026
3.5×
price gap for the same H100 GPU — $1.99/hr on RunPod vs $6.88/hr AWS on-demand. Most startups have never benchmarked their rate.
Published rates, verified May–Jul 2026 · full table below
For every $1M in AI product revenue, ~$230K goes out as inference cost (ICONIQ State of AI 2026) The directional 2026 Series A bar: $3.5M ARR · 120%+ NRR · 60%+ gross margin (Carta · CRV) 73% of organisations blew through their AI budget (FinOps Foundation 2026) AI contract renewals reportedly coming back at 4–5× — vendors hold every comp; buyers hold one H100: $6.88/hr on AWS · $1.99/hr on RunPod — same chip Over 60¢ of every VC dollar went to AI companies in Q1 2026 — the highest share on record (Carta) Blended token costs fell 67% in a year — AI bills kept climbing anyway (2.4B API calls) AWS raised reserved GPU prices twice in 2026 — +15% Jan · +20% July (SemiAnalysis · The Register) CME, ICE & AX have all announced compute futures — compute is becoming an asset class For every $1M in AI product revenue, ~$230K goes out as inference cost (ICONIQ State of AI 2026) The directional 2026 Series A bar: $3.5M ARR · 120%+ NRR · 60%+ gross margin (Carta · CRV) 73% of organisations blew through their AI budget (FinOps Foundation 2026) AI contract renewals reportedly coming back at 4–5× — vendors hold every comp; buyers hold one H100: $6.88/hr on AWS · $1.99/hr on RunPod — same chip Over 60¢ of every VC dollar went to AI companies in Q1 2026 — the highest share on record (Carta) Blended token costs fell 67% in a year — AI bills kept climbing anyway (2.4B API calls) AWS raised reserved GPU prices twice in 2026 — +15% Jan · +20% July (SemiAnalysis · The Register) CME, ICE & AX have all announced compute futures — compute is becoming an asset class
When this becomes urgent

Three moments decide your compute economics.
All three have a clock on them.

Trigger 01 · The cliff
Your cloud credits are expiring
AWS, Google, and Azure startup programs hand out five-to-six-figure credit packages — and the first unsubsidized bill is where margin stories die. The 90 days before expiry is your maximum negotiating leverage. After, you're a price-taker.
Benchmark before the cliff →
Trigger 02 · The renewal
A renewal just landed in your inbox
Renewals reportedly coming back at 4–5× are priced against your ignorance: the vendor holds every comp they've ever signed, you hold one. Send it to us before you reply to it.
Send us the renewal →
Trigger 03 · The raise
You're raising in the next 12 months
Compute unit-economics questions are now asked at every round, and the 60%+ gross-margin bar is graded on your signed contracts, not your projections. Walk into diligence with renegotiated terms.
Fix the margin story →
The gap

Your inference bill decides your next round.
And no one negotiates it for you.

The 2026 fundraising benchmarks are explicit: Series A investors now cite 60%+ gross margins as the bar, while the average AI product builder runs 52% — with inference alone averaging roughly 23% of revenue. Your compute contracts are no longer an infrastructure detail. They are the difference between a fundable margin story and a pass.

Now look at who could help — honestly. Cloud cost consultancies exist, but the leading ones brand themselves for companies spending $100M+/yr, and even the ones without a stated floor negotiate the AWS bill and stop there. Commitment-automation tools optimize reservations inside one cloud. Your fractional CFO firm builds excellent models of the problem — but has never sat across the table from AWS, CoreWeave, or Anthropic. And your dashboards just show you the bill.

None of them — not one — negotiates the contract that is actually eating your margin: the model API commit. Your OpenAI, Anthropic, and Bedrock agreements are the fastest-growing line on your P&L and the least-negotiated contracts in your company.

The precise hole in the market. If you're Seed to Series C and spending $20K–$500K a month across model APIs and GPUs, nobody negotiates your inference contracts, and nobody anywhere connects those contracts to the fundraise. Ferrous does exactly that pairing: startup-scale AI contracts, negotiated with enterprise-grade rigor, wired directly into the margin story your investors grade.
Buyer-side only — and it matters. Your cloud provider profits when you spend more. Your tool vendors profit when you keep the stack. Resellers earn margin on what you buy. Ferrous is paid by you and nobody else — the only party at your table whose economics allow the sentences that save you money: cancel the commitment, switch the provider, kill the workload.
~$230K
Inference cost per $1M of AI revenue.
That's the average — before negotiation, routing discipline, or caching. Every point recovered goes straight to the gross margin investors are grading.
ICONIQ State of AI · Jan 2026 (inference ≈23% of revenue)
$0
What your compute costs today — if you're on credits.
Startup programs from AWS, Google, and Azure defer the bill, not the problem. The contracts you sign while spending free money set the rates you'll pay when it runs out. Negotiate before the cliff, not after.
AWS Activate · Google for Startups · Microsoft for Startups
2×
AWS raised reserved GPU prices twice in 2026.
+15% in January — the first increase in EC2's twenty-year history — then +20% more on July 1, across six GPU instance families. AWS's own pricing page says rates are "updated periodically based on supply and demand." Unnegotiated buyers are price-takers by default.
The Register · AWS pricing docs · The Information · Jan–Jul 2026
ProviderH100 /GPU-hrType
Azure NC H100 v5$6.98Hyperscaler
AWS P5$6.88Hyperscaler
GCP A3 High~$3.00Hyperscaler
Lambda Labs$3.29–3.99Neocloud
Spheron$2.54Neocloud
RunPod community$1.99–2.39Neocloud
Thunder Compute$1.38Neocloud
AWS savings plan$1.90–2.10Reserved

Representative on-demand rates as published May–July 2026 · Spheron · IntuitionLabs · Cast AI · ThunderCompute · Lambda. Actual pricing varies by region, configuration, and pricing model (on-demand vs capacity blocks vs reserved) — which is exactly why benchmarking your specific contract matters.

This table is negotiation ammunition — and the market just split. Frontier-grade committed capacity is repricing sharply upward while commodity on-demand stays flat: which tranche your contract sits in now decides everything. Whether prices fall (we arbitrage the spread) or rise (we lock your rate and capacity before the repricing), the vendor across the table knows every price they've ever charged. Every benchmark memo we run adds startup-scale deal terms to the dataset — increasingly, so do we.
What we do

We start with a free benchmark.
We end at your term sheet.

Founding cohort — savings-share only. Our first client cohort pays nothing fixed: we take a share of verified savings, and only after they land. If we don't move your numbers, you owe us nothing. In exchange, your (anonymized) deal terms seed the startup-scale benchmark dataset every future client benefits from.
01
The entry point · Free · 48 hours · No call required
The benchmark memo
Send one AI or compute contract, renewal quote, or invoice. Within 48 hours you get a two-page memo — yours to keep whether or not we ever speak again. It's the fastest way to find out if you're overpaying, and the only cost is an email.
What you pay vs. marketYour effective rate against verified published rates and startup-scale comps for the same capacity.
Your three leverage pointsThe specific terms — commit size, tranche, escalators, capacity language — where you hold cards you haven't played.
The number to ask forWhat this contract should cost, and the realistic range a negotiation lands in.
Capacity security gradeWhether the contract guarantees capacity or just price — because in this market, price without capacity is half a contract.
02
The leverage layer · Savings-share pricing
Negotiation & migration
Then we run the negotiation with you. Model API commits — OpenAI, Anthropic, Bedrock, Gemini, Mistral — sized to validated base load, not hope, with capacity reservations to match. Hyperscaler agreements with multi-cloud optionality as live leverage. Rate-locks and capacity guarantees timed ahead of the next repricing. Provider migration to Lambda, RunPod, or Spheron at $1.99–3.99/GPU-hr versus $6.88 where workloads allow — with availability risk priced in, because the cheapest rate is worthless if the capacity isn't there at renewal.
03
The judgment layer
Spend-to-value verdicts
Negotiation fixes the price; judgment fixes the quantity. We connect verified cost data to output and deliver kill / keep / constrain decisions on every AI workload. Model routing and caching architecture worth up to 8× on blended token cost — which then becomes leverage in step 02, because a credible plan to route work off frontier models changes what commit size you should sign. Agent governance — step budgets, context discipline — for the least self-limiting spend in your company. Tools detect anomalies. We make rulings.
04
The fundability layer
Compute unit economics your investors will fund
The raise is where it all lands. We rebuild your financial model with compute as the variable cost it actually is — by GPU class, provider, and workload — and build the gross-margin trajectory from the ~52% industry average toward the 60%+ investors now benchmark against, backed by contracts we actually moved. Board packs, chargeback governance for token-based spend, venture debt structured against compute commitments, and the answers ready for the compute questions now asked at every round.
What the benchmark surfaces
01
Renewals priced against no benchmark
Vendors quoting 4–5× at renewal because the buyer has never seen a comparable deal. The single fastest-payback finding we make.
02
Hyperscaler rates at 3× market
AWS and Azure for workloads that could run on Lambda or Spheron at a fraction of the cost.
03
Commitments sized to hope
Reserved capacity and API commits bought against projections, not validated base load — 30–50% billed in full and unused.
04
Frontier models for routine work
Every workload routed to the most expensive model. Tiered routing cuts blended token cost by up to 8× — $2.31 vs $18.40 per million.
05
Agents without budgets
Autonomous workflows retrying failed calls, re-reading the same files, running forty doomed steps. The fastest-growing spend and the least self-limiting.
06
Duplicate API subscriptions
Multiple teams paying separately for OpenAI, Anthropic, Bedrock — often for overlapping capabilities.
07
GPU for CPU-eligible workloads
Embedding generation and preprocessing running on GPU when CPU costs one-tenth as much.
08
Training costs misclassified
R&D training runs billed as COGS — quietly wrecking the gross margin number your investors are grading.
For investors & platform teams

You set the 60% bar.
We get your portfolio over it.

You're the one citing the gross-margin benchmark in partner meetings — and the one watching portfolio companies miss it because of contracts nobody at the company has ever negotiated. Every point of inference cost recovered is a point of gross margin at the next round, and a cleaner markup for your fund.

Ferrous runs a portfolio compute screen: a benchmark memo on the top compute contracts across your Seed-to-C AI companies, delivered to you and each founder. You see which margin stories are at risk before diligence finds them. Founders get the numbers and the leverage. No cost to the fund.

What the portfolio screen delivers. A one-page flag report for the fund — which companies sit above market rate, which have renewals inside 6 months, which have credits expiring — plus an individual benchmark memo for each founder who opts in. Anonymized, confidential, and yours within two weeks of the contracts arriving.
Why funds go first. The compute questions are already in your diligence templates. Getting the answers fixed before the round — rather than discounting for them at term sheet — is cheaper for everyone at the table.
What makes Ferrous different

Three differentiators.
Only one is permanent.

01 —
Buyer-side only
Permanent
No vendor relationships, no referral fees, no reseller margin, no commissions. We can tell you AWS is 3× market. We can tell you to cancel a commitment. We can tell you the free tool is enough. Our revenue does not depend on your spend continuing — which is precisely why our advice is worth paying for.
02 —
The contracts no one else touches
The pairing no one else offers
Cloud consultancies negotiate the AWS bill — at enterprise scale. Automation tools optimize reservations inside one cloud. CFO firms model the margin, never move it. Nobody negotiates model API commits at startup scale, and nobody connects any of it to the fundraise. Ferrous does both: the inference contracts and the unit-economics story they have to add up to. Your investors get a narrative backed by signed terms, not assumptions.
03 —
Startup-scale deal comps
Compounds with every memo
Every verified rate, contract term, and renewal delta at startup scale is logged anonymously — from every benchmark memo, free or paid. Enterprise firms hold enterprise comps; no one collects the $20K–$500K/month deal data — the deals your vendors price against your ignorance. We publish the aggregate quarterly as the Startup Compute Price Index, so the dataset is verifiable, not a claim. This is the moat.
The forward view

Compute is becoming an asset class.
The comps decide who benefits.

The advisory we provide today delivers immediate, measurable value without requiring any financial market to exist. The renegotiated contracts, the recovered margin, the funded rounds — none of this depends on regulatory approval or market liquidity.

But the longer-term direction is clear: CME, ICE, and AX have all announced compute futures products, and Silicon Data's GPU Forward Curve now offers a standardised 12-month view of anticipated GPU rental costs. When hedging instruments arrive for compute, the firms holding the best transaction-level comps will price them best — and the companies whose contracts built those comps will be first to benefit.

The sequence. Negotiate the contracts. Verify the savings. Publish the index. When the instruments arrive, Ferrous will be the only firm that has done all three for this segment — and its clients will be holding contracts structured to hedge.
CME Group ICE / NYSE AX Exchange Ornn / OCPI Silicon Data / DRW GPU Forward Curve
Oct
25
October 2025
Ornn OCPI goes live (per Ornn). First transaction-based GPU pricing benchmark — printed trades, not surveys.
Dec
25
December 2025
First compute swap executes via Ornn Exchange (per Ornn).
Jan
26
January 2026
Architect (AX) & Ornn launch exchange-traded perpetual futures on GPU and RAM prices.
May
5
May 5, 2026
Larry Fink at Milken Institute points to compute futures as an emerging asset class — Bloomberg.
May
12
May 12, 2026
CME Group & Silicon Data announce compute futures (DRW-backed).pending approval
May
18
May 18, 2026
ICE (NYSE owner) & Ornn announce cash-settled GPU compute futures covering multiple GPU types.pending approval
Start here

Send one contract.
Get numbers in 48 hours.

One AI or compute contract, renewal quote, or invoice — that's all the free benchmark needs. Two pages back within 48 hours: what you pay vs. market, your leverage points, and the number to ask for. Yours to keep. No call, no proposal deck, no commitment.

Paid by you. Only you.
Founding cohort pays only from verified savings. No referral fees, no commissions, no vendor relationships. We can tell you to cancel — and we do, when the numbers warrant it.
The benchmark stands alone
48 hours, free, no retainer. You get real numbers on your actual contract whether or not we ever work together.
Confidential by default
Contracts are seen by Ferrous only. Deal terms enter the benchmark dataset anonymized and aggregated — never attributable to you.

We respond within one business day. All information is confidential.

Message received. We will reply within one business day with a secure way to send your contract — your benchmark memo follows within 48 hours of receiving it.
Ferrous
ferrous.pro For venture-backed AI companies Data sourced & checked July 2026 Get your free benchmark ↗