Nvidia disclosed this week that its supply commitments have reached $279 billion — contracts to buy components from its own suppliers, signed years ahead of the revenue those parts will eventually produce. Strip the zeros off and it is the same decision a four-person company makes when a vendor offers 22% off for a three-year term.
You are being offered money to give up flexibility. The only question that matters is whether you are being offered enough.
Most people answer by staring at the discount. That is backwards. The discount is the price. What you need first is the value of the thing you are selling.
Step 1: Price the option you are selling
When you sign a multi-year commitment, you sell the vendor an option: the right to keep your money regardless of what happens to your business, their product, or the market rate. That option has a value, and you can estimate it in about fifteen minutes.
Take four numbers:
- C — the annual spend you are being asked to commit
- N — the number of years
- D — the discount offered, as a percentage
- P — your honest probability that you would want out before the term ends
Sign only if D × N × C is comfortably greater than P × (the cost of being stuck).
The cost of being stuck is not the remaining contract value. It is the remaining contract value plus what you would pay for the replacement while still paying for the thing you have stopped using. Double-running is the line everybody forgets, and it is usually the larger number.
The arithmetic, once
$60,000 a year, three years, 22% off. You think there is a 30% chance you want out in year two.
- Discount saved: 0.22 × $60,000 × 3 = $39,600
- Cost if stuck: ~$60,000 of remaining commitment plus ~$50,000 of replacement running in parallel = $110,000
- Expected cost: 0.30 × $110,000 = $33,000
- Margin in favour of signing: $6,600 on $180,000 of spend — an edge of 3.7%
A 22% discount sounds enormous. Against a one-in-three chance of needing out, it is close to a coin flip. That is the whole point of running the numbers: the headline percentage is designed to feel decisive, and it almost never is.
The break-even shape, memorised
For a three-year term with realistic double-running, the discount you need just to break even runs roughly like this:
- 10% chance you want out → you need about 6% off
- 20% → about 12%
- 30% → about 18%
- 40% → about 25%
- 50% → about 33%
Rough, but directionally right. If a vendor offers 15% and you know in your gut there is a one-in-three chance you will want out, the deal is bad, and no amount of relationship warmth changes the arithmetic.
Step 2: Three questions that decide it
The maths gives you a number. These three questions tell you whether to trust it.
1. Is this a commodity or a bet?
Commodity: electricity, bandwidth, warehouse space, payroll software, anything whose price has been boring for three years and whose substitutes are obvious. Commit freely.
Bet: anything where the category itself may look different in twenty-four months. New tooling categories. Model providers. Anything whose list price has moved more than 30% in the last year in either direction. Do not commit.
The tell: if the vendor’s own roadmap could make your commitment obsolete, you are on the wrong side of the trade. A vendor who plans to launch a cheaper tier next year has every reason to lock you into this year’s pricing now.
2. Does your usage curve go up, or sideways?
Commitments are safe when you are growing into them and dangerous when you are buying your current peak. Vendors price the deal off your present usage and let you feel shrewd for locking it in.
The rule: commit to 60–70% of forecast usage, never 100%. Buy the floor, pay list on the spike. An occasional expensive spike costs less than a permanently stranded floor, and it keeps you honest about what you actually consume.
3. Who owns this in year two?
The most common failure is administrative, not strategic. The person who negotiated the deal leaves. Nobody remembers the terms. Minimums go unmet, credits expire quietly, the contract auto-renews on a date nobody diarised.
If you cannot name the human who owns this contract and point at the calendar reminder set for 90 days before renewal, you are not ready to sign. That is not a metaphor — make the reminder before you countersign.
Step 3: The commitment ladder
Five structures, safest first. Work up this ladder before you accept the headline discount.
- Month-to-month, list price. Zero commitment, highest unit cost. The baseline everything else is measured against.
- Annual term, monthly billing. You commit to twelve months; you pay monthly. Most of the discount, none of the cash risk. This is the sweet spot for most buyers, and most vendors will grant it if asked plainly.
- Annual prepay. Pay up front for another 5–10%. Worth it only if the cash is genuinely idle and the vendor is financially solid enough that you would lend them the money unsecured. Because you are.
- Multi-year term with an annual ramp. Year one at 60% of the eventual number, stepping up. Legitimately good if you are scaling and can defend the forecast.
- Multi-year prepay with a use-it-or-lose-it pool. Maximum discount, maximum danger. Unused credits expire. This is where cloud commitments and AI credit deals live, and where the money quietly dies.
Most buyers leap from rung one to rung five, because that is where the big number is printed. Rung two typically captures around 70% of the savings for a small fraction of the risk.
Step 4: Four clauses that make a lock-in survivable
Negotiate these before the price. Once you have agreed the number, you have spent your leverage and the vendor knows it.
Rollover
Unused commitment carries forward instead of evaporating.
“We’re comfortable with the volume. We’re not comfortable with a hard expiry on unused credits. Can unused balance roll into the following term, capped at 25% of the annual commitment?”
Downward flex
One chance per term to reduce the commitment by a defined percentage.
“We’d like a single one-time reduction right — up to 20% of the annual minimum, exercisable once, 60 days’ notice. That’s the thing that lets us sign three years instead of one.”
Price protection
If list price falls, or a cheaper tier appears that fits you better, your rate follows it down.
“If you release a lower-priced tier that covers our use case during the term, we want the right to move to it. We’re not asking for a refund — just not to be the last customers paying last year’s price.”
Termination for degradation
Not termination for convenience — vendors will refuse that and you will burn goodwill asking. Tie exit to something objective: a sustained SLA breach, a discontinued capability you depend on, a material change in terms.
“If you deprecate a capability we’ve named in the order form, we need the right to exit without penalty. Tell us where to name it and we’ll sign.”
This is the clause most people fold on, and it is the one that matters most. Naming your dependency in the contract converts a vague fear into a defined trigger — and it forces a useful internal conversation about what you are actually depending on.
Our take: The discount is never the deal. The deal is the discount minus the option you just wrote, and the option is priced in flexibility you will not miss until the quarter you need it. Ask for the clauses first, the price second, and treat any vendor who will not discuss structure as one who is counting on you not to.
Worked example: the three-year cloud contract
Current spend: $18,000 a month, growing about 4% a quarter. The vendor offers 30% off a three-year commitment at $20,000 a month of committed spend, billed monthly, unused capacity expiring each month.
Run the check. The commitment is above current usage — you are buying your forecast, not your floor. Growth is 4% a quarter, so you reach $20,000 in roughly six to seven quarters. That is five or six quarters of paying for capacity you do not use, at up to $2,000 a month of waste.
Counter. Commit at $12,000 a month, roughly 65% of current usage, and accept a smaller discount on the committed portion. Consumption above the floor pays the on-demand rate. Total saving drops from a theoretical $216,000 to about $130,000 — but the $130,000 is real, and the $216,000 was always partly fictional.
Then add the clauses. Monthly rollover of unused committed spend within each quarter. A one-time 20% step-down at the end of year one. Price protection against new instance types. Sign that.
Worked example: the AI credits prepay
A model provider offers 25% off if you prepay $100,000 of credits, valid twelve months, unused balance forfeited.
Run the check. Question one already fails: this is a bet, not a commodity. Per-token pricing in this category has moved sharply and repeatedly, and every provider ships cheaper tiers on a cadence measured in months. A 25% discount against a category that has repriced by more than that inside a year is not a discount — it is a bet that prices stop moving.
The right answer is usually no. If you sign anyway, sign for six months, not twelve, and insist on price protection so a mid-term price cut flows through to your remaining balance. A twelve-month forfeiting prepay in a category that reprices every quarter is the single worst structure on the ladder.
Failure modes to recognise
- The peak-usage trap. Committing to your busiest month and paying for it in your quietest. Always commit to the floor.
- The orphan contract. Nobody owns it after the champion leaves. It renews itself for years.
- The discount stack. A vendor combines term, prepay and volume discounts into one number so you cannot see which lever cost you what. Ask them to price each separately.
- The relationship discount. A better rate offered informally, on a handshake, not written into the order form. It does not exist.
- The sunk-cost renewal. Renewing a commitment you never fully used, because switching now means “wasting” what you already spent. That money is gone either way.
- Signing at quarter end. Their urgency, not yours. It is real leverage — use it for structure, not just price.
The twenty-minute version
- Write down C, N, D and an honest P. Compute both sides. If your edge is under 10% of total spend, the deal is not worth the rigidity.
- Ask: commodity or bet? If bet, stop. Take the annual term instead.
- Cut the committed volume to 60–70% of forecast usage.
- Ask for rollover, downward flex, price protection and termination-for-degradation — in that order, before discussing price.
- Name the internal owner. Set the 90-days-before-renewal reminder now.
- Then, and only then, negotiate the number.
Nvidia can carry $279 billion of committed spend because it has visibility into demand that almost nobody else has, and because being rationed on memory would cost it more than being over-committed. That is a real calculation, made with real information. The version of it landing in your inbox this week comes with a countdown timer and a 22% badge, and the person who sent it has better information about their roadmap than you do.
Do the arithmetic anyway. It takes fifteen minutes and it is the highest-paid quarter-hour in your month.