Sablewick — AI Inference Infrastructure Exercise
An 8-hour Venture Investment Exercise case study with a complete model answer
Modeled After
Greenoaks
The open exercise format Greenoaks is reported to set, applied to an infrastructure business: long-term margin structure, what the company does to its industry, and a valuation from scratch — with the margin answer feeding directly into the terminal assumptions of the model.
Structure and exercise format are modeled after Greenoaks — the take-home exercise format the firm is reported to use. The company, the financials and every figure in this case are entirely our own.
The Situation
Sablewick Compute, Inc. runs a GPU cloud and an inference API.
Sablewick Compute, Inc.
- Sector
- AI infrastructure — a GPU cloud and inference API, selling serverless inference by the token, reserved training capacity by contract, and dedicated deployments to customers whose fine-tuned weights live on the fleet
- Size
- Geography
- United States, with the fleet across leased data center space and an installed cost of $43,200 per accelerator including its share of the network fabric, the storage tier and the power and cooling plant
- Ownership
- Situation
The Prompt
You are a candidate for an investment role at a fund that holds positions for years rather than quarters. The exercise arrives by email on a Friday afternoon with a short note:
"Sablewick.
Supporting Materials
What you are handed at the start of the case, in the format a real process would use.
Blank modeling template (template.xlsx)
The price panel (given in the template)
Market and third-party data (given in the template)
What You Have to Produce
The deliverables, in the order the committee will read them. The exercise runs 480 minutes.
PART 1
Pick a unit, and pick it so the exhibit adds up
PART 2
Find both plugs, and notice they are different lines
PART 3
Treat the depreciation schedule as a judgment, not an input
PART 4
Print the margin on both bases, every time
PART 5
Model revenue as price times volume, moving in opposite directions
PART 6
Separate the three revenue lines, and price the concentration
PART 7
Build the fleet forward, and let capital expenditure fall out of it
PART 8
Model buy versus lease; do not assert one
PART 9
Sensitize on the judgment, not on the discount rate
PART 10
Answer the industry question with a counterfactual
Attempt It First
Blank modelling template
The answer model with every produced cell cleared — the shell you build your attempt in. Work it in Excel against the clock, then check yourself against the model answer below.
How to Approach It
The order a strong candidate works in, and why. This is the shape of the answer — the finished memo and Excel model are in the solution set below.
- 0:00 – 0:45
read everything, and decide about the panel early
- 0:45 – 2:15
per-accelerator economics
- 2:15 – 3:45
revenue and token economics
- 3:45 – 4:15
buy versus lease
- 4:15 – 5:45
the discounted cash flow
- 5:45 – 6:00
sensitivities
- 6:00 – 8:00
write it
Key Concepts
The ideas this case is built on. Know these cold and the case becomes a question of execution rather than knowledge.
Price per million tokens
The unit an inference API sells in, and the number that falls fastest in this industry. Revenue is price times volume, and modeling them as one line hides the question: volume can grow enormously while revenue barely moves, and the year those two curves cross is the most important year in the forecast.
Revenue per deployed accelerator-year
What one machine earns in a year. It is the right unit for this business because it is the only one on which the capital cost of the asset and the operating cost of running it are the same size — and because throughput improving while price falls faster is what makes it decline even as volume rises.
Economic versus accounting useful life
The life on the financial statements is a policy; the life over which an asset actually earns is a judgment. A generation of silicon that resets the price of compute can end the economic life of the generation before it while the older machine still works perfectly. In a business this capital-intensive that judgment is worth more than every revenue assumption combined.
Cash gross margin versus margin after depreciation
The first is the measure the market quotes and it is a software measure. The second charges the asset that produced the revenue. Quoting the first alone is the easiest way to overstate a hardware business, and it is the thing a reader will check first.
The plug
The line in a bridge that is not assumed but falls out of the others as a residual. Whichever line is the plug is the claim the write-up has to defend, and an exhibit that does not name its plug is presenting an assumption as an output.
Quality of revenue
Not all revenue is worth the same multiple. Revenue sold on multi-quarter contracts to a handful of counterparties who could build the same capacity themselves is worth less than revenue from thousands of customers whose weights live on your fleet — and a blended net revenue retention figure will conceal exactly that difference.
Equivalent annual cost
The level annual charge with the same present value as an irregular one. It is what makes owning an asset and renting it comparable on one line, and it is the only honest way to compare a capital outlay recovered over years against a rent paid every year.
The value-driver identity
Terminal free cash flow expressed as after-tax profit less the capital that growth consumes, where the capital growth consumes is the growth rate divided by the return earned on it. For a business whose growth requires buying depreciating hardware, capitalizing the last explicit cash flow instead would price a moment in a build cycle rather than a steady state.
Degenerate sensitivity axes
Two drivers that enter a model only through their sum or their product are one driver. A grid built on such a pair is constant along its anti-diagonals: it looks like a two-way sensitivity and carries one dimension of information. Testing one off-diagonal pair before shipping a grid costs nothing and catches it.
What Makes It Hard
The specific traps in this case — the places candidates lose the assessment without noticing.
Check Your Answer
Type the figures you produced and find out how many are right before you open the worked answer. You get a verdict and, where you are off, a pointer to the part of the build to re-check — never the number itself. Everything you type stays on this device.
How to type a figure. Digits, with an optional unit: 1,234.5, $1,234.5, 2.6x, 21.4%. For a negative use (20.0) or -20.0. Enter as many decimals as you carried — precision is never penalised.
What the Case Asked For
The arithmetic is only half of it. Tick off what you actually produced — this half is yours to score, because nothing can grade a written recommendation from a checkbox.
The Model Answer
The worked answer in full: memo and Excel model, built the way a banker would actually build them. It is a reference, not a submission — a strong answer under the clock is far shorter.
What the Solution Covers
- —Token economics and price per million tokens
- —Hardware capex and depreciation exposure
- —Inference versus training versus fine-tuning revenue mix
- —Customer concentration and quality of revenue
- —Pricing compression underwriting
- —Buy versus lease capacity
Memo
The written recommendation and how it was reached
Upgrade to Diamond
Sign up and upgrade to Diamond to unlock the memo, the Excel model and the audio walkthrough.
Get StartedThe Excel Model
The model is linked and tied out end to end — every schedule, every formula and every check, in the file itself.
Downloads are available to Diamond members
Excel Model and Memo (PDF) — yours to open, edit and rebuild
Walkthrough
A conversational walkthrough of how to approach the case under time pressure — where to start, what to cut, and how the recommendation gets defended.
Audio Walkthrough
How to approach Sablewick — AI Inference Infrastructure Exercise
60-second preview — upgrade to Diamond for the full walkthrough
Frequently Asked Questions
Is this a margin exercise or a valuation exercise?
Both, and they are the same exercise. The per-accelerator work sets the steady-state dollars; the cash flow model walks today's reported actual to those dollars and prices the result. If your terminal column is not identical to your bridge, you have built two models rather than one, and a reader will find the seam.
What unit should the margin analysis be denominated in?
Something on which the capital cost and the operating cost are the same size. Per million tokens the capital cost is invisible; per dollar of revenue the depreciation charge moves whenever price moves. One deployed accelerator over one year works, and it has a second advantage: if you choose the denominator carefully, the printed dollar column and the printed percentage column both add up exactly as rendered.
How should I handle the depreciable life?
As a judgment, not as an input. Compute the whole answer twice, on two materially different lives, and print both. Then take a position and say what evidence would move you — the utilization of the fleet's oldest cohort and the secondary market for retired accelerators are both observable, which is what makes this a decision you can revisit rather than a guess you are stuck with.
Can I use EBITDA here?
You can compute it, and you should probably quote it once so the reader knows you have seen what the price looks like on that measure. But EBITDA in this business adds back the cost of the asset that produces the revenue, which makes it the least informative of the three multiples you can strike. Saying so is worth more than the multiple is.
What do I do with the price panel?
Work it up, extract the one defensible reading, name in specific terms what you could not resolve about it, and then put it in an appendix and wire it into nothing. A list price is not a realized price, so it can describe the direction of price and not its level — and a valuation needs the level. The temptation is either to ignore it because it is awkward or to lean on it because it was given to you. Both are wrong.
Is there a leveraged buyout in this case?
No, and importing one is the most common way to get this exercise wrong in kind rather than in detail. There is no sponsor, no acquisition debt, no debt schedule, no cash sweep, no exit multiple and no returns bridge. The equipment-backed borrowing on the balance sheet is one line in an enterprise-value bridge and the reason the buy-versus-lease question has a balance-sheet answer as well as a margin answer.
Should I value a security or a business?
The prompt asks what the business is worth, so give a whole-company value — and then say out loud that a whole-company value is not the value of any one class, because the preference stack was not supplied. Naming that limitation is worth more than assuming it away, and it is one of the few places in this exercise where the honest answer is that you do not have enough to answer.
Is this format still used?
Yes, and it is the most transferable of the long-only and growth formats because it is closest to the actual work. A firm sends three open questions and one imperfect file, and wants to see how you decide what is worth knowing, where you stop, and what you are willing to put your name to.
About This Venture Investment Exercise Case Study
Venture Investment Exercise case study for venture & growth interviews. 480-minute format covering token economics and price per million tokens, hardware capex and depreciation exposure, inference versus training versus fine-tuning revenue mix. Includes the full prompt, a written memo, a tied-out Excel model and an audio walkthrough.
This case study sits in Venture & Growth, under Investment Exercises. Every case ships with the full prompt, the supporting materials, a complete model answer and an audio walkthrough of the judgment behind the recommendation.
480-Minute Format
The time limit a real assessment would give you
Memo
Included in the model answer
Excel Model
Included in the model answer
Audio Walkthrough
How to approach the case under time pressure
200K+ students have used IB Vine to help land offers at top firms.
Moelis & Co.
“I loved IB Vine. I’m going into Private Capital Advisory (PCA); you have the lessons on there for PCA and that was actually a game changer. I went into those interviews and knew about the core PCA concepts like net asset value, primary fundraising vs. secondary advisory, etc. I swear I was on IB Vine like a few hours a day. The opportunity to have that learning section and go through the questions people submit and everything — there’s no better tool out there. I loved the guides/lessons. The user interface is amazing. The way you’ve simplified it is so spectacular — it’s just so much easier to digest. It’s fun to use too; I’d rather use IB Vine than scroll through a guide. To be honest, I think the product is perfect. I genuinely owe you a big thanks.”
Moelis & Co.
“IB Vine is a tool we really love to use in the club I’m a part of, and there’s really no other resource like it. You guys do a phenomenal job with the question bank. I recruited specifically for Energy banking, and IB Vine was my most used resource for generalist questions (which were about 50% of my interview questions; the rest being Energy-specific); the majority of such questions I saw in interviews were at least similar (if not the same) to the ones on IB Vine.”
Perella Weinberg
“Once I read through the BIWS learning guides, I really didn’t refer to them again. I didn’t even really run through the 400 question guide once I found IB Vine, which I heard about through one of my classmates. We even get a free subscription (like most business schools) to Wall Street Prep, and if I’m being completely honest, I never even logged in to WSP. IB Vine is pretty much the only tool I used (along with our club question bank & mock interviews with peers) and it was invaluable for recruiting.”
Cantor Fitzgerald
“What a platform, made such a huge difference. I did superdays at Evercore, PJT, M. Klein and Barclays among others from a non-target school and did not miss a single technical in any interview process through prepping with IB Vine.”
Barclays
“Very accurate questions and all of the solutions are easy to follow. At least 10 of the questions I studied through this platform appeared in my Round 1 or Superday interviews.”
TD Securities
“THANK YOU SO MUCH IB VINE, I COULD NOT HAVE DONE THIS WITHOUT YOU, SERIOUSLY!!! IB Vine was the best website ever. I spent at least two hours on this daily (seriously) from October through I get my offer in February.”
Citadel
“I’m doing public equities this summer and next. I know the name is “IB Vine” but at the undergrad level a lot of the technicals across public equities are the same as investment banking. I recommend your software to all my friends!”
Piper Sandler
“I loved this site! 1000% this is the best resource I used in the process.”
Lazard
“This was the greatest tool ever. I genuinely enjoyed running through the technicals/behaviorals and it was very helpful!”
Houlihan Lokey
“IB Vine was the most helpful resource I had during recruiting. I will continue to promote it to other students at WashU and elsewhere.”
Jefferies
“Awesome product, helped me crush my technicals in my interviews and land a great role. Thank you, seriously was a huge help.”
Morgan Stanley
“Very helpful to get real-life questions unlike the 400 guide, especially for merger math. Built deeper understanding of key concepts.”
Houlihan Lokey
“Extremely helpful study tool that carried me through the recruitment process from start to end.”
Evercore
“You guys are doing great work over there with IB Vine. Absolute staple for interview prep.”
Rothschild
“IB Vine was incredibly helpful and I am forever thankful for all the help.”
RBC
“IB Mock was amazing - I used it for multiple hours. Also the flashcards and the lessons features on IB Vine were fantastic. Thank you!”
Dragoneer
“IB Vine is such a great platform, really impressive. There is so much value in this. The audio podcasts / mock interview library are pretty incredible.”
Case Study Preparation
Explore All Case Studies
100+ case studies, each with the full prompt, supporting materials and an audio walkthrough; most also ship a model answer deck and a tied-out Excel model.
Every case has a public page like this one. The member library is the signed-in index members work through.