GAVIN BAKER: A contract is a promise and a vague memorandum of understanding is not, so only contracted commitments should be counted

The Gist

Gavin Baker draws a line between signed contracts and loose memos. Contracts count; memos that either side can walk away from do not. Several big 2025 AI announcements were memos. This is a steelman: each side is reconstructed in its strongest fair form, not endorsed, and no winner is declared.

Conclusion

Commitment totals should count only contracted or funded obligations; non-binding memoranda inflate headline figures and do not represent promises to pay.

Premises

1. Long-term agreements, supply capacity agreements, take-or-pay contracts and externally funded data center projects are contracts, and contracts are promises that counterparties take seriously.
2. A memorandum of understanding with many outs for both parties creates no enforceable obligation.
3. Several large AI supply announcements in 2025, including memoranda with Asian memory suppliers, were of this non-binding type.
4. If such memoranda become contracts, they will be counted.

Assumptions

• Baker does not quantify how much of the roughly $1 trillion consists of memoranda as opposed to contracts.
• Nvidia's $100 billion OpenAI letter of intent was never finalized and was replaced by a roughly $30 billion equity investment in 2026.
• OpenAI's 900,000 wafer per month memory target with Samsung and SK Hynix remained a non-binding letter of intent with minimal 2026 volumes as of October 2026.

Analysis

Argument type: Deductive.

Potential Fallacies

• Hasty generalization / selective examples (Premise 3 and Conclusion) — A few prominent non-binding deals (the Nvidia-OpenAI letter, the Samsung and SK Hynix letters) are used to support a claim about the whole headline total. Because the memorandum share is unquantified, the examples show that some inflation exists but not how much. They are also the most salient cases, so they may not be representative.
• False dichotomy (binary treatment of a graded matter) (Premises 1–2 and Conclusion) — Commitments are sorted into binding promises and non-promises. In practice, credibility is graded. Some contracts have broad exit rights or weak counterparties, and some non-binding agreements carry real strategic and reputational weight or contain partly binding terms such as exclusivity. Probability-weighted or tiered reporting would track this better.
• Equivocation on 'commitment' / 'promise' (Premise 1 and Conclusion) — The argument moves between legal enforceability, the counterparty's seriousness and the likelihood of eventual payment. Memoranda fail the first test, but that does not settle the third. Contracts pass the first without guaranteeing the third.
• Unsupported equation of contract status with reliability (Premise 1) — Premise 1 treats contracts as promises 'counterparties take seriously' without evidence on termination, force majeure, renegotiation or counterparty credit. Contract status is an imperfect proxy for realized spending, especially with cash-burning buyers or circular vendor financing.
• Missing normative bridge / is-ought gap (Premises 1–3 to Conclusion) — The premises describe legal status, while the conclusion prescribes an accounting standard. The needed principle, that commitment totals should track enforceability, and the factual claim that current headline totals include memoranda are both left unstated. P3 shows memoranda existed, not that they were counted.
• Persuasive definition and loaded framing (Title and Premise 2) — Terms such as 'vague memorandum', 'many outs' and 'inflate' build the conclusion into the description of the category. This treats every memorandum as the weakest kind and implies deliberate padding.

Counterarguments

These challenges are not ranked.

• Premise 1 (High impact) — Legal enforceability is not performance. Take-or-pay and capacity contracts are often renegotiated, deferred or defaulted on, and a contract with a cash-burning or highly leveraged counterparty may be weaker than a non-binding commitment from a well-capitalized, motivated one. Telecom and fiber contracts in 2000–2001 are a historical example. Circular financing, where vendors fund customers who then buy from them, can make even contracted totals soft and correlated.
• Conclusion (High impact) — A binary in-or-out rule is both over- and under-inclusive. Probability-weighted or tiered reporting (binding, funded, non-binding with milestones, aspirational) would preserve information about intent while still separating promises from hopes. Excluding memoranda entirely discards demand signals that shape supplier capex and capacity planning.
• Conclusion (High impact) — The claim that headline figures are 'inflated' is unquantified. If memoranda are a small share of the roughly $1T, the correction is immaterial. If a large share, the correction is significant but the argument does not show it. Without a decomposition, the inflation claim is directional only.
• Premise 3 (Medium impact) — The examples are selected, salient failures. Without base rates for how often letters of intent convert into contracts in comparable supply deals, two or three cases cannot show that non-binding items are typically hollow. Hindsight also makes ex ante signaling value look smaller than it was.
• Conclusion (Medium impact) — Headline totals may serve different purposes: measuring legal exposure, forecasting demand, or signaling ambition. A legal-enforceability test suits only the first. Announcers may also have presented figures as 'up to' or aspirational, in which case the criticism misreads what the figure claimed to measure.
• Premise 4 (Medium impact) — A contract-only rule invites Goodhart effects: parties may paper memoranda as thin contracts with generous exits so they qualify for the count. It also makes the total a lagging indicator, and it does not say how partial conversions (such as $100B becoming $30B of equity) should be handled.
• Premise 2 (Low impact) — Preliminary agreements are not uniformly unenforceable. Some impose binding duties to negotiate in good faith, exclusivity or reliance obligations, and repeat players in tightly interdependent supply chains may honor letters of intent for reputational reasons.

Suggested Improvements

• Quantify the composition of the total — Decompose the roughly $1 trillion into binding contracts, funded projects, equity-linked deals and non-binding letters or memoranda, and state the non-binding share. This would turn a directional point into a demonstrable materiality claim and answer the main empirical gap.
• Replace the binary with tiers — Propose labeled categories such as binding take-or-pay, binding with outs, funded, non-binding intent and aspirational, each reported separately, or probability-weighted using observed conversion rates. This retains the honesty goal while capturing signaling value and avoiding both over- and under-counting.
• Address contract quality — Acknowledge that contracts vary in firmness and require disclosure of termination rights, take-or-pay floors, counterparty credit and whether funding is vendor-linked. This closes the gap between 'enforceable' and 'reliable' and preempts the strongest counterargument.
• Define 'funded' and 'contracted' — Specify whether 'funded' means cash committed, project financing in place or equity invested, and set a bright-line test for 'contracted' (for example, firm minimums net of termination rights). The disjunction currently blends different kinds of assurance and could be gamed.
• Supply base rates and broader evidence — Include historical conversion rates for letters of intent in comparable deals, along with examples of both honored memoranda and cancelled contracts. This would address cherry-picking and show whether the cited cases are typical.
• Specify the purpose of the metric — State which question the count answers, whether legal exposure, expected spend or demand signal, and apply the standard symmetrically to bullish and bearish claims. The right counting rule depends on purpose, and stating it makes the prescription more defensible.

Scenario Tests

• Nvidia's $100B letter of intent is never finalized and is replaced by a roughly $30B equity investment. (Supports) — The headline figure overstated what materialized, which supports the claim that non-binding announcements can inflate totals. It is a single case, and the equity investment, which may be vendor-financed, shows that 'funded' can be circular.
• OpenAI's memory letters with Samsung and SK Hynix remain non-binding with minimal 2026 volumes. (Supports) — The 900,000 wafers per month target had not been backed by firm volumes as of October 2026. Ramp timing could explain part of this, so it supports the point more weakly than the Nvidia case.
• A take-or-pay contract with a cash-burning AI lab is renegotiated or deferred when its funding tightens. (Challenges) — Contract status did not translate into payment, so the contract/memorandum line is an imperfect quality filter. A contracted-only total could still overstate firm demand.
• A non-binding letter of intent between two strongly interdependent firms is honored: the supplier builds capacity and volumes ramp as planned. (Challenges) — Excluding the letter would have understated real commitment, which favors tiered or probability-weighted counting over a binary cut.
• A decomposition shows that non-binding items make up a small fraction of the roughly $1T. (Challenges) — The normative rule stays valid but the claim of material inflation weakens, reducing the argument to a minor clarification.
• A memorandum converts into a signed contract and is then added to the total. (Neutral) — This is consistent with P4 and shows the rule is dynamic. It says nothing about how common such conversions are, or whether the contract carries firm obligations.

Coherence & Relevance

The core logic is coherent: if a total is meant to represent promises to pay, counting non-enforceable memoranda overstates it, and P1–P2 form a clear dichotomy. The argument is incomplete because two bridging premises are unstated: that commitment totals should track enforceability, and that current headline totals actually include memoranda at face value. The inflation claim is limited by the unquantified memorandum share, and the contract/non-contract binary is cruder than the real range of commitment strength. The result is a well-grounded definitional and disclosure point that is stronger as a call for transparent categorization than as a demonstration that the headline figure is materially overstated.

• Long-term agreements, supply capacity agreements, take-or-pay contracts and externally funded data center projects are contracts, and contracts are promises that counterparties take seriously. (Strong) — Defines the category to be counted but equates legal status with reliability and treats 'funded' and 'contracted' as interchangeable. It does not address counterparty credit, exit clauses or circular financing.
• A memorandum of understanding with many outs for both parties creates no enforceable obligation. (Strong) — Establishes the exclusion criterion, but does not show that non-enforceable means uninformative or unlikely to be performed. It also assumes every memorandum fits the 'many outs' description.
• Several large AI supply announcements in 2025, including memoranda with Asian memory suppliers, were of this non-binding type. (Moderate) — Shows that non-binding items exist but not that headline totals include them at full value or what share of the roughly $1T they represent.
• If such memoranda become contracts, they will be counted. (Weak) — Shows the standard is even-handed and dynamic, but does not support the claim that current totals are inflated. It leaves conversion verification and partial conversion unaddressed.

View this argument on LogicFirst.ai