Should You Buy Tokens Forward? How It Paid Off Big Time for This Customer Support Company
Most companies treat inference as a cost to manage. Clearpath treated it as a lever. Here's what that looked like six months later.
Clearpath and all figures in this article are illustrative. The framework and tradeoffs are real.
How fast the price of compute can move
In the four weeks from December 9, 2025 to January 6, 2026, the hourly rate to rent an H100 GPU jumped 10%, from $2.00 to $2.20, according to Silicon Data's GPU Rental Index.
Token forwards are one answer to that. They let companies procure inference tokens in advance, lock in unit economics for the full term, and secure supply before demand outpaces available capacity. The buyer commits to a volume, the provider guarantees both the price and the allocation. But a token forward can be a blessing to your business or a trainwreck. Most companies get it wrong, but the ones that get it right are ripe to leapfrog their competitors.
To show you how to get it right, we'll walk through a worked example: a customer support company called Clearpath that was once in your shoes. We'll cover how they thought through it, the questions they asked, and what actually happened six months later. The company and figures are illustrative, but the framework and the tradeoffs are real.
Clearpath builds customer support software
Their AI assistant handles tier-1 tickets (password resets, order lookups, basic troubleshooting), and it's working. Response times are solid, the engineering team knows the integration, and customers aren't complaining.
Then they heard about a fast growing competitor that was using token forwards
Clearpath's competitor had made a move. They'd locked pricing before a rate increase and secured guaranteed capacity before the market tightened. Clearpath was still paying as they went. One of Clearpath's engineers had seen it come up in a Slack community they both followed. It wasn't a rumor. The competitor wasn't just ahead on cost; they'd stepped out of the on-demand queue entirely. Clearpath's team started looking into what that actually meant for their own situation. They didn't want to fall behind during one of the most important periods in their company, let alone in history.
What everyone gets wrong about token forwards
A token forward isn't a volume discount. It's a bilateral commitment contract, standardized and traded through a marketplace. The buyer commits to a specific volume of inference expressed in Standardized Token Units (STUs); the provider locks in the per-unit rate and priority allocation for the full term. The key difference from negotiating directly with a provider: STUs are fungible across verified providers in the marketplace, so offers are genuinely comparable, and committed traffic gets priority that on-demand traffic doesn't.
One scope constraint: token forwards apply to open-weight models. Llama, DeepSeek, Qwen class, and similar. Closed-API providers have their own volume commitment programs with different mechanics. Clearpath ran on open-weight models, so token forwards applied directly.
The opportunity was real. But so was the risk of getting it wrong.
Clearpath's expected inference usage over the next six months would cost $100,000 at current on-demand rates. A token forward would let them lock in $80,000 for that same allocation, a 20% reduction. Beyond the price, there was a supply angle too: demand for open-weight models had been climbing, and locking in capacity ahead of potential availability constraints was worth something on its own. But their product team was also in the middle of rewriting their prompt templates, building a retrieval layer to cut context length, and evaluating a smaller model that might handle half their ticket types just as well. Customer volume was growing, but the amount of inference they'd actually need was genuinely unclear.
Rather than jump at the headline number, Clearpath's engineering lead sat down with finance and worked through four questions before anyone signed anything.
- How much usage can we confidently commit to?
- Does the discount still save money if demand disappoints?
- What happens if we change models or the provider has problems?
- What do we need to bring to a quote request?
But first, they stepped back to understand this key context
Before evaluating the numbers, Clearpath's team made sure they understood what they were actually agreeing to. A token forward, as offered through verified token forward marketplaces (more on that shortly), works in three stages: Commit. Lock. Tap. The buyer specifies their volume and term. The provider locks in a per-unit rate and priority allocation for the full term. The buyer then draws down against that committed balance as they use it, against their real demand curve.
The thing that makes this different from just negotiating a discount: volume is committed in Standardized Token Units (STUs), a common unit across providers. You're not holding a provider-specific credit that locks you into one relationship. You're holding a standardized, fungible unit that any verified provider in the marketplace can fulfill against. That makes offers genuinely comparable, and it means the priority allocation is real. On-demand traffic can get deprioritized during demand spikes; committed STU balances don't.
Locking in a rate makes their inference budget more predictable, and the priority allocation protects them during high-demand periods. But it doesn't guarantee the locked rate will be cheaper than what they could buy on-demand six months from now. And committing to an allocation doesn't, by itself, promise anything about latency during their Monday morning rush. Throughput targets need their own agreed terms.
For the math below: the $80,000 obligation covers a defined six-month allocation, unused volume expires without refund, and rates are fixed. Rollover terms, settlement schedule, and overage pricing each vary by provider, and those details shaped the final decision as much as the headline discount.
There were three real options on the table
The first thing Clearpath's team agreed on: don't let the new offer become the only option on the table. Their current provider had earned some goodwill. The engineering team had spent months tuning that integration, and they knew exactly how it behaved when ticket volume spiked on Friday afternoons. Any competing offer needed to justify the cost of replacing that.
They laid out three real paths:
| Path | What they needed to evaluate |
|---|---|
| Stay on demand with current provider | Keep full flexibility. No minimum commitment. Their actual negotiated rates were already below list price. |
| Negotiate a commitment with current provider | Ask whether volume commitment earns better pricing, without the risk of migrating anything. |
| Evaluate token forward offers through a marketplace | See if a standardized forward contract with a competing provider improved the total deal enough to justify evaluation work and potential migration. They used Compute Exchange, a marketplace that aggregates committed-use quotes across verified open-model providers. |
The comparison had to be against what they could actually get, not list prices. Their current provider already gave them a volume discount. A competitor's headline 20% reduction might only be 10% better in practice.
Compute Exchange made the third path easier. Clearpath could run that comparison through a marketplace instead of cold-calling providers one by one. Same workload, same quality requirements, same service expectations, across all three options side by side.
Now the big question 1: how much usage could they actually commit to?
The $100,000 forecast was built on assumptions: customer growth, average conversation length, and how many tickets the AI handles versus escalates. The team's first move was figuring out which parts of that number were solid and which were optimistic.
The honest answer: a lot of it was optimistic. Their engineering team was actively working on three things that would each reduce token usage: shorter prompt templates, a retrieval system that pulled in only the relevant parts of a customer's order history instead of the whole thing, and a smaller model for simple ticket types. If those projects landed over the next six months, token volume could drop significantly even as customer volume grew.
They worked through a concrete version of that tension: ticket volume might grow 25%, but average tokens per conversation could fall 20% from the prompt and retrieval work. Net result: usage stays almost flat.
1.25 × 0.80 = 1.00
Total token volume stays essentially unchanged despite growth. That only holds if the efficiency improvements actually ship on schedule, and if the token mix (input vs. output vs. cached-input) stays roughly similar.
Their traces were the key input here. Pulling 90 days of inference data, they split tickets by type: straightforward ones (password resets, order status) versus complex ones (refund disputes, multi-step troubleshooting). The complex tickets were using roughly 8x the tokens of simple ones. If the smaller model could handle the simple category, the committed baseline dropped meaningfully.
They also ran evals on the simplified prompts before treating that reduction as bankable. A prompt that uses 30% fewer tokens but causes enough quality degradation to increase retries doesn't actually save anything.
The output of that work: two forecasts. The growth-plus-efficiency case. And a conservative case: what does usage look like if the retrieval system ships two months late and the model evaluation drags? They decided to commit against the conservative number, not the optimistic one.
They also confirmed that Compute Exchange's forward structure uses a single six-month STU consumption window with no monthly minimums, not $13,000 worth of STUs per month with expiring buckets. That mattered a lot for their seasonal ticket patterns.
Question 2: Would the discount hold if things didn't go to plan?
The offer looked like a $20,000 saving. But Clearpath knew that number was conditional. It only materialized if they actually used what they committed to. So they stress-tested it.
| What could happen | On-demand cost | Committed cost | Difference |
|---|---|---|---|
| Usage hits full forecast: growth comes in, efficiency projects slip | $100,000 | $80,000 | Save $20,000 |
| Usage comes in at 80%: one efficiency project ships early | $80,000 | $80,000 | Break even |
| Usage drops to 60%: retrieval layer works better than expected | $60,000 | $80,000 | Pay $20,000 more |
Break-even was 80% consumption. If their efficiency work succeeded beyond the conservative forecast, the commitment would cost them money. The "20% discount" was really a bet that usage stayed high.
The "20% discount" was really a bet that usage stayed high. Get the committed volume wrong and you're not saving money, you're paying for tokens you don't use.
They also tested a market-rate scenario: what if on-demand prices dropped 20% industry-wide during the commitment? It's happened before. At 80% usage, they'd be paying $16,000 more than going month-to-month. That was the downside they were accepting.
Their conclusion: the bet was reasonable, but only if they committed at the conservative usage number, not the optimistic forecast. They also factored in the evaluation time for a new provider. If that cost $5,000 in engineering hours, break-even moved from 80% to 85%. They built that into the decision.
Question 3: What if things changed mid-commitment?
This was a factor that almost killed the question internally. Clearpath's model evaluation was actively in progress, and there was a real chance they'd want to shift 40% of simple tickets to a smaller model three months in. Would a commitment accommodate that, or would they be locked into paying for tokens on a model they'd partially moved off of?
They went through every plausible change scenario and made sure the contract addressed each one before signing:
| What might change | What they nailed down before signing |
|---|---|
| Efficiency work cuts usage below commitment | Confirmed: unused STUs roll over within the six-month window. No monthly minimums. |
| The smaller model passes evals and they shift volume | Negotiated model portability: remaining STU entitlement can move to the new model at an adjusted rate. |
| The provider has a rough stretch | SLAs defined with exit rights, not just credits. If availability falls below threshold for two consecutive weeks, they can walk. |
| Ticket volume surges past the allocation | Overage rates pre-negotiated at the same per-STU price as the commitment. No separate agreement needed. |
They also ran the prospective provider on a real sample of their tickets before signing: not a demo, but actual production-representative queries, including the ones that had historically caused retries. The provider passed. A support assistant that generates a wrong answer and has to try again isn't just using double the tokens. It's the kind of experience that ends up in a customer complaint.
Getting those contract terms took one round of negotiation. The provider agreed to all of them. That's when Clearpath felt confident enough to sign.
So they decided to request a quote. Here's how they nailed it.
When Clearpath went back to the market (through Compute Exchange, alongside their current provider), they came prepared. The quality of the brief directly affects the quality of the offer. They brought six things:
- Usage history: 90 days of monthly input, cached-input, and output tokens, separated by type, with daily peaks flagged so providers could see the Friday afternoon pattern.
- Two forecasts: the growth scenario and the conservative scenario, with the specific assumptions behind each one made explicit.
- Model requirements: their current model and version, plus the smaller model under evaluation, with an explicit ask for STU portability if they moved traffic.
- Service requirements: response-time targets, peak traffic windows, data handling requirements, and SLA thresholds they needed to see in the contract.
- Eval set: 75 real tickets covering the full complexity range. This is what they used to test every provider before including them in the comparison.
- Commercial baseline: their actual current rates, not the list price. Any offer had to beat what they could already get.
They asked every provider for two quotes: one at the conservative forecast, one at the full forecast. The spread between those two numbers told them a lot about how each provider was pricing the risk: flat discount or volume-sensitive pricing. That shaped the negotiation.
What actually ended up happening
After working through the four questions, Clearpath made a call: commit to a six-month forward, but sized to their conservative forecast, not the optimistic one. The final commitment was $72,000, not $80,000. Smaller than the original offer, better contract terms, and a number the engineering lead, finance team, and product manager could all sign off on.
The savings were real. But the moment that mattered happened mid-term, when Clearpath's competitor started throttling.
Six months later (illustrative figures)
| Metric | What happened |
|---|---|
| Actual usage | Came in at 91% of the conservative forecast. The retrieval system shipped on time and cut context length significantly. The smaller model is still in evaluation. They didn't need to use the STU portability clause, but they're glad they negotiated it in. |
| Cost vs. on-demand | Saved $14,200 compared to what the same usage would have cost at on-demand rates. Less than the headline $20,000, but a real, defensible number based on what they actually consumed. |
| Capacity and supply | Mid-term, open-weight model capacity tightened across several providers. Clearpath's committed STU allocation held. In a scenario like this, a competitor running on-demand would likely face degraded response times and throttling during the same period. Clearpath's committed allocation would insulate them from that. That capacity guarantee isn't in the $14,200 savings figure, but it's the kind of operational headroom that's hard to price until you need it. |
| Engineering impact | The retrieval layer shipped on schedule, partly because the team wasn't scrambling to manage a blown inference budget. Budget certainty created room to actually execute the efficiency roadmap rather than react to it. |
| What they'd do differently | Commit sooner. The time they spent evaluating a new provider (which they ultimately didn't switch to) cost about three weeks of engineering time they could have skipped by negotiating harder with their existing provider first. The process for the next renewal is already mapped out. |
Capacity held
Priority allocation
When open-weight supply tightened mid-term, Clearpath's committed STU balance kept their queue position. On-demand traffic gets deprioritized during spikes. Committed doesn't.
Budget locked
Fixed unit economics
A known per-token rate for the full six months meant no mid-quarter surprises. The engineering team built and shipped the retrieval layer without scrambling around a moving inference budget.
Cost saved
$14,200
Clearpath paid $72,000 for inference that would have cost $86,200 at on-demand rates. Less than the headline $20,000, but based on what they actually consumed.
The number that matters isn't the advertised discount. It's the amount you can confidently commit to, the contract terms that protect you if things change, and the capacity guarantee that holds when on-demand doesn't.
The lesson Clearpath took from it: the number that matters isn't the advertised discount. It's the amount you can confidently commit to, the STU structure that maps honestly to your actual token mix, and the terms that protect you if things change. Get those right, and a token forward isn't a bet. It's a straightforward win.
The companies that get this right don't look back
The Clearpath example points at something replicable. The savings were real but secondary. What mattered more was that they treated inference as a financial input to actively manage rather than a variable cost to react to. That shift changes how the whole evaluation works: you're not just asking "can we afford this," you're asking "what do we need from the contract to make the economics work?"
That means a conservative committed volume, contract terms that protect you if usage shifts mid-term, and a proper apples-to-apples comparison against what you can already get. Done right, a token forward isn't a bet on usage staying high. It's a structured hedge with a known downside.
If you want to see what an offer looks like for your workload, Compute Exchange aggregates quotes across verified open-model providers in Standardized Token Units, so the comparison is direct.
The company, offers, and calculations in this guide are hypothetical and provided for illustration only. They are not Compute Exchange quotes, standard contract terms, or Silicon Data price forecasts. Verify all terms directly with any provider before making a commitment.