When consumers are asked to pay for something, they think twice about whether to consume the offered product / service or not.
There’s a cognitive overhead in all decisions but the one involved in micropayments is acute because it’s more mentally taxing to decide whether or not to spend, say, $0.10 to read an article than the $0.10 itself. Ergo, as Byrne Hobart notes in The Diff Capital Gains newsletter, consumers’ subconscious veers towards the all-you-can-eat subscriptions, which spare them the cognitive overhead of the buy-or-not-buy decision. That’s also probably why Why Publishers Don’t Sell Individual Articles even though micropayments have been around for a decade or longer.
There’s a similar cognitive overhead in B2B technology procurement.
Back in the day, when companies learned that their servers and storage had an average utilization of only 30-40%, they were attracted by the usage-based pricing model offered by public cloud infrastructure providers like AWS and Azure. For somewhat similar reasons, they moved from onprem COTS enterprise applications to SAAS software. However, when they’re asked to pay $100/month to add another user for the SAAS software, CIOs wonder if an existing user can’t handle the extra load and CFOs question if it’s necesary to buy an additional user subscription.
In addition, cloud creates the chore / anxiety / terror (choose one depending on company culture) of IT having to go to finance every month to get vendor payments released on time. God save the CIO if, in a given month, the cloud service provider slaps her with overage charges and the bill doubles or triples (which happens more often than you might think, see example here).
All of these tasks create massive cognitive overhead for CIOs.
To sidestep it, many CIOs sign up for three year subscriptions and pay upfront for one year or more – which defeats the “pay-per-use” and “cancel anytime” raison d’être of cloud computing. (This is not the same as Cloud Repatriation or SAAS Repatriation wherein enterprises move their workloads totally from public cloud back to onprem.)
Coming to accounting, when a customer signs up for three years and pays the entire subscription fees upfront, the cloud service provider / SAAS vendor still cannot bill the total contract value (TCV) at one shot. This is because GAAP revenue recognition norms allow billing only in proportion to value rendered. Since value is rendered continuously throughout the entire contract period, billing and revenue recognition can happen only on a monthly or max quarterly basis.
Due to a similar rule for cost recognition on the buyer’s side, the aforementioned customer can’t book the cost of 36 months subscription fees at once but can do so only on a month-by-month basis. This explains how a huge cash outflow can coexist with a small cost.
It could also shed light on the DeepSeek controversy.
While most of us may have heard of DeepSeek only a week or two ago, Matt Levine wrote about it in his Money Stuff column back in June 2024.
Chinese quant hedge fund Zhejiang High-Flyer Capital Management (“High Flyer” from here on) recently launched an Artificial Intelligence Large Language Model called R1. DeepSeek sent shockwaves throughout the world for three reasons:
- It matched the performance of the most powerful reasoning model in the world, namely, OpenAI o1.
- Its training cost was only $6 million (as against the hundreds of millions spent by OpenAI, Anthropic and other western AI firms).
- It was open-sourced.
On the very first day of trading after DeepSeek open-sourced R1, stocks of semiconductor and electric utilities crashed by a trillion dollars, with $NVDA alone suffering a $600 billion loss of market cap.
With due respect to the engineering innovation involved in R1’s high performance, the aforementioned financial accounting jugglery could explain the low training costs and open-sourcing of DeepSeek.
While DeepSeek’s claimed performance is verifiable, there’s tons of skepticism over its claim of getting R1 trained for a mere $6M (which surely played a role in its owner’s ability and decision to open-source it).
David Sacks, the new AI Czar of USA, asserted that that DeepSeek spent over a billion dollars to buy an NVIDIA GPU cluster.
@DavidSacks: New report by leading semiconductor analyst Dylan Patel shows that DeepSeek spent over $1 billion on its compute cluster. The widely reported $6M number is highly misleading, as it excludes capex and R&D, and at best describes the cost of the final training run only.
Even if we take Sacks’s assertion at face value, it’s quite possible that DeepSeek booked only the OPEX cost for the actual hours of training DeepSeek, which would obviously be a tiny fraction of the $1B CAPEX figure.
I was sold on this argument after reading Ben Thompson / Stratechery‘s take on this subject:
DeepSeek claimed the model training took 2,788 thousand H800 GPU hours, which, at a cost of $2/GPU hour, comes out to a mere $5.576 million.
This should explain the disconnect between $1B and $6M and settle the controversy over DeepSeek’s training cost.
Switching between CAPEX and OPEX is similar to the Sale and Lease Back model used in the commercial real estate industry. Under SLB, a company sells a property it owns (e.g. office building, warehouse, or retail space) to an investor or real estate firm, leases back the property from the new owner under a long-term lease agreement, and continues to use the property while paying monthly or periodic lease payments. Click here to find out the benefits of SLB to both the original owner and the new owner.
Going by the number of conspiracy theories that High Flyer may have shorted Nvidia, Broadcom, other semi and electric utility stocks and pocketed a fortune, it can’t be ruled out that DeepSeek spun a high CAPEX investment as a low OPEX spend to create the trillion dollar market meltown last week.
What are the chances that @deepseek_ai’s hedge fund affiliate made a fortune yesterday with short-dated puts on @nvidia, power companies, etc.? A fortune could have been made.
— Bill Ackman (@BillAckman) January 28, 2025
We may never know whether High Flyer did any such thing but Reuters reported that some others did short the Nvidia stock and rake in over $6 billion in profits after DeepSeek panic.
