The workload can move before the bill falls

1. The answer

Technically feasible substitution is already an operating reality for the selected workloads. What is not established is a universal cost advantage, an immediate reduction in the buyer's bill, or the disappearance of NVIDIA's pricing power. The most consequential new distinction is between keeping a customer's cloud spending and keeping that customer's workload on a particular chip. A portable spending commitment can accomplish the former while permitting the latter to change.

The new comparison follows an actually deployed image-generation service, a separate real-time camera application, the cloud contracts under which such services can be bought, and the resources required to keep supplying them. It identifies three economically different outcomes. An unconstrained buyer can retain a component saving after the remaining service and switching costs. A buyer with an unused, non-portable commitment can pay more by switching too early. A cloud operator can preserve a committed receipt while directing the work to its own accelerator. None of those outcomes alone identifies the chip manufacturer's retained profit.

The quantitative core starts with Loka's reported 17.5% lower cost per generated image for fine-tuned SDXL on Inferentia2 against its optimized NVIDIA-based G5 comparison. That is a historical implementer measurement, not a current all-in saving. Under an explicitly illustrative one-year allocation of the reported ten-million-image projection, three months of stranded old-family payments turn the $14,353 component advantage into a $6,151 additional cost, before switching and other differences. Waiting for the old commitment to expire—or moving within an eligible portable commitment under the stated no-other-use assumptions—retains $10,765 of the reference advantage. These are calculated decision conditions, not observed customer bills.[1]

The threshold is the useful result: with that cost ratio and no offsetting use, only 2.1 months of stranded payments in a twelve-month comparison exhaust the nominal annual component saving. A lower billed rate therefore does not settle the economically relevant decision. The calculation also identifies the alternative to a forced stay-or-switch choice: wait, then switch. It does not classify Loka's client as locked into any particular contract.

The strongest favorable interpretation for the incumbent remains substantial. Compatible software, lower latency, an already paid commitment, ready capacity and a smaller validated fleet can justify a premium even when another machine has a lower active-work cost. The strongest adverse interpretation does not require AI demand to vanish: a working alternative can capture suitable tasks, or strengthen bargaining, while total useful computation grows. The new evidence puts more weight on contestable allocation of a real and growing market, not on either an empty market or guaranteed permanent supplier rents.

For the whole bubble question, this is a narrower but consequential advance. The productive-service case becomes stronger; a simple inference from more useful compute to unchanged incumbent margins becomes weaker. The observations do not establish the prevalence of overvaluation, an industry loss total, or that NVIDIA's costed favorable ownership path in “Ownership prices and accumulated economics” is impossible. They establish additional conditions that such a path must satisfy: retained architecture share, effective prices, paid system renewal and customer demand that can actually use the capacity.[18][19]

2. What the deployments can establish

“Second-cycle recovery” established that older designs and facilities can continue to earn. “Efficiency, competition and full-cost returns” investigated accepted-output costs using a bounded OCR service comparison. “Ownership prices and accumulated economics” valued a profitable supplier while charging research, compensation and renewal; it did not establish the durability of every competitive advantage. “Customer funding and repeat investment demand” investigated customer accounts, funding and further purchases. Those findings remain distinct from the service-level comparison here.[15][17][18][19]

This chapter changes the unit of investigation to a deployed service that can move between architectures. Its new quantitative contribution is a joint account of switching, contractual cash floors and finite replica capacity, followed by a supplier-side investment reconciliation. It does not reuse the old A100/H100 tariff ratio, OCR remaining-cost budget, NVIDIA supply-payment scenario or parent funding sensitivity as the new result.

2.1 The principal deployment: a matched workload, with a bounded measurement

Loka's 18 December 2024 account describes a client deployment of fine-tuned SDXL 1.0 for interior-design images. It compares g5.8xlarge with inf2.8xlarge, optimizing the former with torch.compile and the latter through Neuron/Torch XLA. Its method uses 25 prompts, five warm-ups and 50 timed repetitions per prompt per platform. It reports the alternative taking 0.31 seconds longer per image, alongside the lower cost, and extrapolates savings to ten million images. The client is unnamed. The reported volume is a projection, not a ledger of accepted designs or paid demand.[1]

The case is preferable to a peak-compute ranking because the application and deployed model family are identified, both implementations were optimized, and the claimed gain concerns a purchased computing service. It is not independent replication. The author supplies implementation expertise and has a commercial incentive to promote the deployment route.

The public text was inspectable; the linked cost/time figures did not yield usable images in the cited source inspection. Consequently, this report does not invent absolute mean times, percentiles, hourly prices or visual quality scores from those figures. A linked public compilation example is not the private fine-tuned checkpoint or an invoice. The arithmetic below uses only the numerical statements actually available in the body. Their rounding and “up to” language limit precision.

2.2 The production contrast: an always-on service, not another image-rate table

The May 2026 AWS/Tomofun account describes moving BLIP inference for Furbo pet-camera behavior alerts from GPU-based on-demand deployments to inf2.xlarge. A Tomofun staff machine-learning engineer is a co-author. The report attributes an 83% deployment-cost reduction to the change, with the core model logic retained behind Neuron-compatible wrappers. The prior GPU model, observation interval and complete cost ledger are not disclosed.[2]

This is evidence of an actual operating alternative, rather than just a demonstration that the code compiles. It is not a valid way to replace Loka's measured ratio with a much more favorable one. Different workloads, GPU baselines, resource sizing and purchasing terms can produce radically different savings without contradiction.

The binding contrast is response service under variable concurrency. Generating a design somewhat later may be acceptable; an alert delivered after its useful moment is not the same product. Tomofun's inspected concurrency chart shows rising latency and saturation as client/server loads change. Its published discussion establishes a need to tune the concurrency range. It does not supply a transferable per-chip fleet capacity or an independently measured tail-latency guarantee. A large connected device base is not the number of simultaneous requests one chip can process.[2]

2.3 The limit of the chosen comparison

The numerical core is a documented within-provider deployment comparison, with the named production migration as a constraint on generalization. It does not pretend to be an independently audited cross-cloud procurement. An owned cluster, a rented instance and a managed API transfer different obligations; only rented instance service is costed directly here.

Amazon/Anthropic's much larger Trainium relationship enters later as evidence of production scale, supply obligations and investment, not as a third performance benchmark. No SDXL or BLIP measurement is transferred to frontier training. AMD and Alphabet were possible acquisition routes, not required company quotas; neither is assigned a margin, market share or presumed technical disadvantage because this comparison is better documented through AWS.

3. Define the work before calling either system cheaper

For this investigation the desired unit is an image or inference result that the application accepts at its required quality and response time, delivered with the necessary availability, data handling and support. A generated image, a floating-point operation or an attached device is not automatically that unit.

The table is a measurement protocol, not six inferred failures. A production operator may already have satisfied these conditions internally. The report cannot import that operator's unpublished conclusion as a universal measured ratio.

3.1 Similar instance names do not mean identical bundles

Current published specifications give both selected eight-xlarge instances 32 vCPUs and 128 GiB of host memory. G5's A10G has 24 GB accelerator memory; Inf2 has 32 GB. G5 includes 900 GB local NVMe storage, whereas Inf2 is EBS-only. Advertised network and EBS limits also differ: G5 lists 25/16 Gbps; Inf2 lists up-to-25/10 Gbps. These are bundle specifications, not measured throughput or proof of interchangeable storage behavior.[3][4]

The economic implications are conditional. A model that does not fit in one device may need partitioning or another instance, changing both communication and cost. More nominal memory is useful only when the workload can use it. A workflow relying on local scratch storage needs a charged and validated substitute; a memory-resident service might not care. Equal host-memory capacity does not prove equal delivered service.

The historical benchmark's same model family and stated optimizations are useful controls. The unreported exact settings still matter. A different image resolution, inference-step count, precision, batch, context or retry policy could change the ratio. For BLIP, a pet-action response and its deadline are the relevant outcome; its language/image inputs cannot be translated into SDXL images or frontier-model tokens.

3.2 Quality-adjusted output is an entry condition, not a convenient penalty

Let a_G and a_I be the fractions of generated results accepted without further work. If active computing cost per generated result is c_G or c_I, the corresponding cost per accepted result is c/a. At the reported reference ratio:

accepted-output component ratio = 0.825 × a_G/a_I.

With all other differences held aside, the alternative must achieve a_I/a_G > 0.825 to remain cheaper. This is an algebraic condition, not an estimate that quality deteriorates by 17.5%. The supplied ledger leaves both acceptance observations blank. The worked cases condition on equivalent accepted output; they do not claim to have measured it.

Retries should enter the generated-work requirement; review labor should enter the service costs. Charging a retry through both a larger work count and a second arbitrary failure surcharge would overstate its burden. Conversely, calling a rejected image “output” understates it. A model that fails a mandatory quality, security, memory or response requirement is outside the economic comparison rather than merely a more expensive version of the same service.

3.3 Reliability has an actual purchased boundary

AWS's Compute SLA distinguishes a 99.5% instance-level commitment from 99.99% region-level treatment for the specified multi-zone deployment. These are defined connectivity/availability and credit arrangements, not guarantees of correct AI answers or the application's end-to-end response time. Credit eligibility, exclusions and non-stacking rules matter; compensation is not a promise to restore the buyer's lost business value.[5]

The economic comparison must therefore fund the required topology, not assume a single cheap instance buys a resilient application. Paid standby is not automatically idle waste: it can buy an output guarantee the application needs. Neither is an arbitrary redundant fleet required for every batch job. The finite-capacity test below makes the selected standby policy explicit and charges it symmetrically.

4. The complete service-cost boundary

4.1 Buying a service is not buying a bare chip

The renter's account is:

complete decision cost = instance/capacity cash + additional service cash + incremental internal work + transition/exit cost − applicable credits or separately demonstrated avoided cash.

The account is prospective at the decision point. A historical implementation payment is not paid again merely because the service continues. Future validation, patches, model adaptation, monitoring and support remain costs. An already paid nonrefundable amount is sunk cash, although abandoning what it buys can require a second service payment. An unpaid binding commitment is a future cash claim. Those cases need different cash presentations even when their incremental decision comparison is similar.

The instance bill already pays the host for a bundle of equipment use, facilities, electricity, network capability and operation. Adding an estimated purchase price for that same host equipment to the renter's bill would double-charge capital. Separately billed storage, transfer, load balancing, monitoring, managed services, support, licenses and applicable taxes still belong in the renter's comparison. Costs can cancel only when the same service and quantity actually remain common to both choices.

A marginal generation bill is consequently an informative core, not the full account. Let M be the additional future transition and net service cost of moving, over the chosen horizon, relative to staying. It can include one-off engineering and the signed difference in ongoing costs. In the worked positive-cost sensitivities M ≥ 0; a documented operating advantage elsewhere may instead reduce the moving cost. The public records do not measure M. It remains a required allowance, not a guessed salary or a permanent “switching tax.”

A buyer already running Inferentia should not be charged the old migration effort again in a keep-versus-renew choice. A prospective new client may have a real incremental migration bill. “Buyer demand and budget substitution”'s distinction between sunk implementation and the next voluntary decision applies here without importing its council fee or labor assumptions.[16]

4.2 Capacity and discounts are different purchases

AWS's published Savings Plans exchange a fixed hourly spending commitment for discounted eligible usage over one or three years. EC2 Instance plans apply within the selected family and region; Compute plans can follow eligible EC2 usage across families and regions. No-upfront, partial-upfront and all-upfront cash schedules differ. A Savings Plan does not itself reserve capacity.[6]

The service terms inspected here were updated 13 October 2026. Sections 5.4.1–5.4.2 preserve existing designated prices against ordinary subsequent price changes and provide for term charges despite termination, subject to stated exceptions. The current FAQ also permits returns for plans meeting all three conditions: an hourly commitment no greater than $100, purchase within seven days, and purchase in the same calendar month; covered usage is then repriced. The model excludes such return-eligible plans and does not claim that every commitment is irrevocable in every circumstance.[7][6]

A reserved capacity slot is a separate matter. AWS bills unused Capacity Reservations; matching running use is not billed twice as both a full unused slot and an instance. Future-dated reservations can carry a minimum commitment period and restrictions on reducing the count or duration. Immediate reservations, future reservations and Capacity Blocks must not be collapsed into one generic cancellation rule.[8][9][7]

These distinctions create a real economic mechanism. A buyer may be able to move code but unable to move a family-specific discount. It may be able to move the discount but unable to obtain the alternative capacity. Or it may obtain both, while an hourly spending floor prevents an immediate cash saving. A public price page alone resolves none of those three cases.

5. The contract-adjusted switching test

5.1 A historical reference, not an invented invoice

Write the reported relative component saving as r = 0.175, the projected dollar saving as S = $14,353, and the projected image count as Q = 10,000,000. If these rounded source statements are applied to the same comparison, the reconstructed cost scaffold is:

G = S/r = $82,017.14 for the NVIDIA-based rental component;

A = G − S = $67,664.14 for the alternative component.

That is approximately $0.008202 versus $0.006766 per generated image. These amounts are derived from the textual projection; they are not separately observed costs. The reconstruction does not recover missing hourly tariffs, actual annual usage or the client's complete costs. The source's “up to” projection and rounded reduction do not warrant dollar-level empirical precision.

For the following decision experiment, distribute the reference work uniformly over twelve equal model months, with the same accepted-output fraction and remaining service requirements. Assume the reference costs represent the effective prices faced in the experiment. This explicitly assumes the historical relative advantage survives whatever plan-specific pricing applies; the actual customer's rate schedule is unknown. The source did not report a twelve-month horizon, a Savings Plan or three months remaining. The comparison is a structural test using historical cost units and currently inspected contract mechanisms, not a historical backtest or a 2026 quotation.

Uniform usage matters because the real commitment is hourly. Unused allowance is not turned into an annual bank that can cover unrelated later peaks. Actual bursty workloads require a time-resolved account and may have less usable coverage than the uniform example.

5.2 Three contracts, and the option to wait

Let f = L/H, the fraction of the decision horizon covered by a remaining commitment. Suppose the old no-upfront, family-specific commitment would cover the old workload's fG cost. There is no other eligible use in the isolated example. M is the net extra future transition/service cost of switching.

Fresh or otherwise fully avoidable purchase:

moving cost = A + M;

saving = G − A − M.

Move immediately while an old-family payment is stranded:

moving cost = fG + A + M;

saving = G − A − fG − M.

This pays the old binding claim during the overlap and the alternative service throughout. It does not call the old payment a second hardware acquisition or an economic loss to every creditor. The old provider may still be paid in full.

Move immediately within a portable commitment:

moving cost = max(fG, fA) + (1−f)A + M.

For a cheaper alternative, the floor still costs fG during the locked interval. Technical work moves; the bill is not refunded in proportion to the smaller use. After expiry, the purchaser can retain the lower avoidable cost, assuming it renews on those terms and capacity remains available.

Wait until expiry and then move:

cost = fG + (1−f)A + M.

With the cheaper alternative and no extra benefit from switching early, waiting and portable immediate migration have the same cash result in this simplified account. Waiting is cheaper than immediate migration with a stranded family-specific payment by fA. This fourth option prevents a misleading binary conclusion that a buyer must either waste the commitment or remain on the old architecture forever.

Calculated, rounded reference amounts. No row is Loka's or Tomofun's observed contract. Additional net transition/service cost is deducted once from the relevant saving. The final two rows coincide only under the stated uniform-cost, no-other-use and no-extra-early-benefit conditions.

For the stranded case, the general break-even condition is:

L/H ≤ (G − A − M)/G.

At M = 0, that is L ≤ 12 × 0.175 = 2.1 model months. Each dollar of additional unavoidable transition cost reduces the allowable overlap. With a shorter intended service life, less of the saving can recover that cost. A long technical equipment life does not establish a long enough contractually usable, demanded workload horizon.

5.3 Reuse, credits and upfront cash change the answer—but only once

Another existing eligible workload can make the abandoned entitlement useful. Let R be the separately demonstrated cash avoided elsewhere by that reuse. The stranded saving becomes G − A − fG − M + R. The companion applies its reuse parameter only to this abandoned-family entitlement; it does not credit the same amount to the separately modeled portable contract. This is not automatically the face value of unused commitment: the avoided alternative bill, rates, eligibility and timing must be identified. Newly invented demand cannot be inserted merely to make the old plan fully utilized.

A portable plan can similarly apply unused allowance to other eligible spending. That may create a real company-wide saving even while the migrating workload's allocated bill looks unchanged. The analysis should use the consolidated purchasing perimeter consistently, not report both a full reduction on the migrated service and a second full reduction on the reused allowance.

A refund, negotiated amendment, promotional credit or valid return may improve the result. The actual amount, expiration and usage restrictions must enter the cash account. A vendor's willingness to discuss a credit is not a received offset; a noncash benefit cannot be used to pay a hardware invoice. The references here establish published mechanisms, not the private customer's negotiated exceptions.

For an all-upfront plan, the earlier payment is not another prospective cash outflow. Staying uses prepaid service; moving may require paying for substitute service during that interval. The incremental overlap can still make early movement unattractive, but the report must distinguish cash already paid from a bill still due. The main table chooses no-upfront terms so its selected future-cash totals are explicit. It does not count an upfront payment and its amortized allocation twice.

A forward-looking decision also needs timing. Migration can begin before a contract expires, with production moved later; an engineering team may validate the alternative without duplicating the full production fleet for the entire interval. Public sources do not price that implementation schedule. The formula therefore does not assign an arbitrary three-month dual-run cost on top of the three-month stranded payment.

5.4 The price concession that changes the buyer's outside option

At a fresh decision, suppose the incumbent rented service cuts its effective price by d, with the same output and availability. The remaining allowance for all incremental switching and service differences is:

M_max = G(1−d) − A = G(0.175−d).

A 10% illustrative concession leaves approximately $6,151 of the reference ten-million-image budget for those differences. A 17.5% concession erases the component advantage before any migration cost. With an illustrative $5,000 net migration/service difference, the old service need cut its bill by about 11.40% to match the alternative's complete modeled cost. None is an observed negotiated concession or a required NVIDIA chip-price cut.

This comparison makes the incumbent's favorable case testable rather than rhetorical. A validated incumbent can retain the work at a premium if the alternative's extra costs exceed the available allowance. Equally, “software lock-in” is not a reason to insert whatever expense would rescue the incumbent. The allowance must be compared with actual avoidable future work, and the alternative must be permitted to win when it fits.

There is actual evidence that NVIDIA-based cloud prices can change: AWS announced June 2025 on-demand reductions of 33% for P4d/P4de, 44% for P5 and 25% for P5en, with a separate effective boundary for new Savings Plans. Those are different instance families from the G5 comparison. They demonstrate an available commercial response, not that G5 received those reductions, that an old plan automatically repriced, or that Inferentia caused NVIDIA's wholesale price to fall.[10]

6. Finite service capacity can reverse the comparison

6.1 More time per output is a capacity requirement, not just a latency footnote

The Loka comparison reports a small time disadvantage for the alternative alongside its lower modeled cost. At ten million generated images, an extra 0.31 seconds per image implies approximately 861 additional serial occupied instance-hours. This is a reconstruction of benchmark-time consumption. It is not measured production utilization, electricity consumption, a requirement for 861 extra calendar hours, or a claim that one user must wait for the entire batch.[1]

The missing absolute time matters. If the GPU takes t seconds per comparable image, the alternative's occupied-time ratio is:

θ = (t + 0.31)/t.

The reported per-image cost ratio s = 0.825 then implies an hourly tariff ratio p_A/p_G = s/θ, only if the underlying cost calculation is hourly price multiplied by comparable occupied time. This is an implied historical price relationship, not a newly observed cloud quotation. It does not reveal the host's cost of making either instance available.

A smoothly divisible workload would need θ times the active resource for the same output. An always-available service instead buys whole replicas. Let x be required load measured in GPU-equivalent active replicas, and z an equal number of paid standby replicas on each architecture. Under the deliberately simplified fixed-latency, equal-quality scheduling assumption:

n_G = ceil(x) + z

n_A = ceil(θx) + z

billed fleet ratio = (n_A/n_G) × s/θ.

This equation pays for the extra instance when the alternative crosses a capacity boundary. It does not assume that the same rented fleet can serve arbitrary additional demand. Standby is charged whether it produces normal traffic or not; it is not automatically free capacity available to rescue a throughput estimate.

6.2 A narrow, explicit example of the capacity step

For illustration only, set the unpublished GPU time to 12 seconds, retain the published 0.31-second difference, and pay for one spare replica on each architecture. These are analytical assumptions, not claims about either customer's production fleet. They give θ ≈ 1.02583 and an implied hourly tariff ratio of approximately 0.80422.

Calculated scheduling illustration, not observed capacity or cost. Equal accepted quality, constant per-image times, no batching changes, one paid spare, and sufficient available instances are assumed. The model does not certify a latency service level or adequate failover design.

Here the alternative crosses the two-active-replica limit at approximately 1.94963 GPU-equivalent replicas. Just below that boundary it uses the same number of rented instances at a lower inferred hourly price. Above the boundary it needs another paid active instance and can cost more despite lower cost in the smoothly divisible benchmark. A different t, workload distribution, standby design or batching implementation moves the threshold. The example's purpose is to identify the observation needed: capacity at the buyer's required service level, not a universal utilization percentage.

Accepted quality changes both sides of the account. If the alternative accepts only a fraction a as many outputs per attempt as the incumbent, it needs θ/a times the active resource for the same accepted output. Its hourly hardware price does not rise merely because more attempts are rejected. The companion keeps those effects separate. Nor does the calculation assume that failed attempts can be rerun until accepted without correlated errors or extra human work.

Tomofun's reported concurrency behavior provides a production-related reason to take this gate seriously: increasing concurrent requests eventually stops increasing useful throughput proportionately. Its public chart does not supply the specific SDXL service-time assumption or the required replica counts above. It is a contrast about the shape of the constraint, not a transferable calibration.[2]

6.3 Cheaper work does not imply unchanged revenue or unchanged investment

At an effective unit price of 82.5% of the old price, a seller needs approximately 21.21% more paid, collected output merely to retain the old rental revenue. That is 1/0.825 − 1, not an estimated demand response. In the 12-second illustration, the active occupied-resource requirement for that output grows approximately 24.34%, before whole-instance rounding, standby and service changes.

These are different quantities. The buyer may choose the same output and keep the saving; buy more output while keeping its budget; or spend less because its need was already met. A provider may improve cost enough that less revenue yields more profit, but that requires a cost observation. Revenue preserved through volume does not by itself preserve operating profit, capital returns or cash after renewal. The next systems must be available and paid for when the capacity limit is reached.

The model therefore neither imposes a universal demand elasticity nor assumes that technical efficiency is demand destruction. The favorable mechanism is real: lower complete cost can make additional useful work worth buying. The unsupported shortcut is to assume precisely the additional paid volume necessary to preserve every incumbent supplier's previous revenue, margin and investment return simultaneously.

7. Supplying the alternative requires investment and enforceable delivery

7.1 A production alternative at scale exists beyond the small comparison

Anthropic's April 2026 account reports more than a million Trainium2 chips already used for training and serving. That is important evidence against treating custom acceleration as merely a demonstration. The associated expansion includes future capacity and future generations; these are not all installed or paid-for output at the announcement date. It does not independently validate the SDXL or BLIP savings, and Trainium is not Inferentia.[11]

Amazon's subsequent June filing records an expansion of the existing AWS commercial arrangement by more than $100 billion over ten years, including performance commitments for AWS chips. The distinction is consequential: the supplier must deliver commercially usable performance, not only reserve a nameplate chip count. An agreement can protect a sales opportunity while imposing a real obligation on the seller.[12]

The funding legs are separate. The filing records a $5 billion Series G investment and another $5 billion Series H investment. The latter reduces a $20 billion financing facility's remaining ceiling to $15 billion. The facility had zero available at inception, with availability tied to delivery milestones; the exact amount subsequently drawable is not established here. Those investments create financial claims. They are not themselves payment for the same amount of computing service, and the ten-year commercial expansion is not a current cash receipt.[12]

This combination supplies a favorable mechanism and a risk allocation. Properly delivered capacity can unlock resources that help a customer use a productive alternative. Performance-linked funding also makes it wrong to treat the entire ceiling as cash available before delivery. Nothing in these records establishes that every invested dollar returned as a cloud payment, that no other customer resources exist, or that financed commerce is fictitious. The new relevance is the linkage between technical delivery and economic availability, not another allegation drawn from an investor relationship.

7.2 The AWS investment account does not permit a costless supply response

The relevant financial perimeter is AWS, not an invented Inferentia product segment. Its reported first-half net property-and-equipment additions accelerated materially alongside revenue and operating income. The following comparison preserves the same six-month periods and uses dollars in billions:[12]

Calculated growth is approximately 32.72% for sales, 41.81% for operating income, and 146.86% for net additions. The operating margin rises from about 36.09% to 38.56%. This is strong evidence that substantial cloud profit and heavy investment can coexist. It does not establish that the new assets have already earned a satisfactory lifetime return, nor that lower Inferentia prices caused the increase in segment margin.

A second, independently useful reconciliation is:

opening AWS net PPE + net additions − reported D&A + unallocated reconciliation = closing AWS net PPE.

Using the reported December 2025 and June 2026 stocks:

190.055 + 90.120 − 15.353 − 1.072 = 263.750 billion dollars.

Thus the stock grows $73.695 billion, compared with $74.767 billion of additions less D&A, leaving a negative $1.072 billion unallocated reconciliation. It is not an identified impairment, a missed supplier payment or a compute-architecture loss. Disposals, allocation boundaries and other movements would need matching before assigning a cause. Reported D&A includes allocated corporate assets, while additions can include noncash items; the figures do not constitute a self-contained cash ledger.[12]

The arithmetic is deliberately not an owner-cash estimate. Subtracting both all additions and D&A from an operating-income number that already includes D&A would charge the same accounting consumption twice. Conversely, treating D&A as the only required investment would erase the documented expansion. A full architecture-specific return requires the actual assets, cash timing, useful lives, utilization, renewal and residual rights. The segment disclosures supply a material resource constraint without supplying that missing allocation.

Power is another supplied-service cost. Amazon describes long-duration energy arrangements with settlement or termination obligations in specified circumstances. These can persist when consumption or deployment changes. They are not free cancellation options just because one architecture uses fewer resources. Neither the entire energy book nor its accounting gains can be assigned to Inferentia. The AWS operating result also contains accounting effects; it is not a clean measure of cash generated by this architecture.[12]

7.3 NVIDIA's response also has commitments and renewal costs

NVIDIA's July 26, 2026 filing identifies $279 billion of manufacturing, supply and capacity commitments, including memory for current and future architectures. The disclosed fiscal buckets sum to that total. Some commitments precede firm customer orders; cancellation, rescheduling or adjustment may entail additional costs. This supports neither “all commitments are freely avoidable” nor “the entire amount is an immediate liability.”[13]

The same filing reports $13.375 billion of first-half research and development expense and approximately $1.6 billion of inventory-only provisions. The broader inventory-and-purchase-obligation provision measure is different; it must not silently replace the inventory-only amount. The provisions are already reflected in the accounts, not an extra future cash loss to subtract again. No source here attributes them to Inferentia competition.[13]

These disclosures change how a supplier-retention case should be stated. Sustaining a competitive system entails paying for design, software, integration and a product transition while managing supply acquired before every final use is certain. A buyer shifting a workload does not automatically cancel the incumbent's earlier procurement. The timing of the effect may first be lower new orders, different system mix, concessions or inventory exposure—not immediate noncollection on every protected sale.

The favorable NVIDIA explanation remains substantial. Integrated systems, networking, compatibility, support and better performance on the buyer's actual mix can make the complete service worth more than a component comparison suggests. A lower competing tariff is not proof that an incumbent architecture loses the workload or must sell below cost. What fails is the assumption that an observed high margin persists without continuing expenditure or a successful response to usable alternatives.

7.4 Where scarce-resource returns can move—and what is not measured

Substitution changes which claims may capture the value. A cloud provider can internalize a portion previously paid to an external accelerator vendor; a software integrator can be paid to make the alternative usable; a customer can retain a lower avoidable bill. At the same time, both architectures need manufacturing, memory, networking and a powered site. A shared bottleneck can limit the speed or reduce the savings of substitution even when the technical alternative works.

The records do not disclose a matched wafer, packaging, memory or complete system bill for these deployments. There is no measured rent transfer to a named foundry or memory supplier in the calculation. Adding their sales to AWS and NVIDIA sales would count intermediate transactions as separate final demand. Likewise, two architectures appearing in a buyer's commitments do not prove two equally incremental spending programmes: one may replace, diversify, insure or temporarily overlap the other.

The material supply conclusion is narrower and stronger: the alternative is demonstrably being deployed, but supplying it at growing scale has asset, performance and funding requirements. Current profitability is favorable evidence; the resource expansion makes free-growth assumptions untenable. The exact division of the resulting return remains unallocated.

8. Who retains the gain?

8.1 The buyer, cloud provider and chip supplier have different accounts

In the avoidable-purchase case, the buyer retains the reduction in its complete bill after incremental internal and transition costs. An integrator receives payment for implementation; that payment is part of the buyer's cost, not a second saving. A lower tariff does not tell us how much of the improvement AWS retains. The host might pass through lower costs, sacrifice margin to attract usage, or improve both its own margin and the customer's economics.

At fixed work quantity, write the cloud provider's full attributable costs as K_G and K_A, including the relevant capacity and continuing service. Its profit difference is:

profit_A − profit_G = (receipt_A − receipt_G) + (K_G − K_A).

At the fresh reference prices, AWS would need to reduce those costs by at least $14,353 over the reference volume merely to preserve the old profit in dollars. That is the reduction in customer receipts, not 17.5% of AWS's unknown cost. Lower prices and higher customer value can coexist with either higher or lower host profit. The missing cost cannot be inferred from the tariff.

With the portable spending floor, the receipt attributable to the locked interval can remain unchanged while the architecture used to satisfy it changes. Any full-cost saving can first benefit the cloud provider, subject to other work, credits and its own supply obligations. This is the chapter's central allocation distinction: a cloud contract can defend the cloud provider's customer relationship without defending the same amount of future demand for an external accelerator.

That does not mean an existing NVIDIA sale is reversed. AWS can continue using already acquired GPUs for other suitable work, retain them as spare capacity, or change future procurement. Each choice has a different cash and ownership consequence. An accounting writedown is not inevitable merely because one workload moved, and an unchanged historical sale does not protect every future sale.

8.2 Growth can share the gain, but the resource bill must grow coherently

For positive per-unit contributions and unchanged fixed costs, preserving contribution dollars after a price/cost change requires:

q_A/q_G = (p_G − k_G)/(p_A − k_A).

This identity is useful only when the per-unit costs are measured on a comparable boundary and both denominators are economically meaningful. Fixed engineering, distribution, standby, new capacity and replacement can also change; they must then be added. The report does not put an invented host cost into the equation to manufacture a favorable or adverse elasticity.

More useful output can be a genuine social or business gain without becoming equivalent new chip revenue. It can also create a profitable new market. The outcome depends on who buys, at what effective price, on which architecture, with what required service and capital. Paying demand—not technical capability alone—must support the additional fleet. The finite-capacity calculation shows why holding all capacity costs fixed while restoring revenue through unlimited volume is inconsistent.

8.3 Test the strongest incumbent narrative on its own economic terms

NVIDIA's October 1, 2026 explanation emphasizes productive systems, a useful installed base and the ability to serve many workloads. The economic core of that argument is credible: redeployability and software continuity can raise utilization and reduce the risk of owning a specialized asset. The article also argues that cheaper computation induces more demand than the efficiency saves. These are supplier arguments, not an independently calibrated demand response for the cases here.[14]

The distinction between owner and renter matters. A cloud operator owning many GPUs may value the option to move capacity across workloads. A customer renting instances for a specific application may instead value the ability to change instance family without owning the idle hardware. The same flexible cloud relationship can therefore preserve the operator's distribution advantage while making architecture substitution easier for the renter.

The new evidence supports part of the incumbent's explanation. Different workloads do require different service envelopes; one platform's versatility can keep hardware useful as applications change. Neither SDXL nor BLIP supplies a reason to assign zero continuation value to the whole installed base. This is consistent with the earlier second-cycle investigation.[15]

It does not establish the stronger universal claim. A system can remain useful while its attainable rental price or utilization falls. More technical output inside a power envelope produces more revenue only when that output is demanded, sold and collected. A diverse workload portfolio can reduce utilization risk without eliminating price competition. And a cheaper alternative can help expand the total service market while taking part of the incumbent's share. These are restrictions on the inference, not a claim that the supplier's engineering improvements are unreal.

No advertised cross-generation throughput multiple, resale-price claim or third-party utilization figure from that narrative is imported into this comparison. Their workload and financial perimeters were not matched to the selected applications. The relevant finding is not which vendor wins every benchmark; it is which combination of useful output, commercial terms and paid-for versatility justifies a premium.

9. What changes in the whole answer

9.1 What changes in the supplier case

The earlier ownership study gave value to a profitable continuing NVIDIA business while paying for manufacturing, research, compensation, maintenance and growth investment. It did not conclude that NVIDIA or Freshworks was demonstrably a bubble. Its stronger cost paths and lower-required-return alternatives remain legitimate countercases. No new quote or stock target is produced here.[18]

This investigation changes the empirical interpretation of four parts of that supplier path:

The 17.5% historical rented-service difference is not a 17.5-point reduction in NVIDIA's manufacturing margin, a forecast annual selling-price decline or an estimate of displaced NVIDIA revenue. The 83% Tomofun claim has an even less comparable financial denominator. Neither can replace a cost package in the earlier ownership model without a matched revenue, bill-of-materials, software and investment account.

Conversely, leaving those quantities unmeasured does not mean competitive evidence is irrelevant. The existence of a production-ready alternative changes the set of feasible buyer choices. The commitment calculation identifies when that choice can alter cash; the supply account identifies what must be funded to provide it. This is new support for a specific vulnerability in rent retention, even though the size and prevalence of an ownership-price error remain unidentified.

9.2 Weigh the favorable and adverse explanations

The strongest favorable explanation is that the market is developing useful services, and both integrated GPUs and custom accelerators can earn by serving different needs. A buyer can retain lower complete costs; a cloud provider can gain scale and better asset use; a GPU supplier can retain valuable work through performance, compatibility and breadth. Actual deployed systems and reported operating profits make that explanation more than a theoretical possibility. Sensible long-term contracts can finance the supply required to serve it.

The adverse explanation need not allege fake demand. The services can work while an investor has assumed too much retained price, share or margin. A customer may migrate at renewal, use a rival only for suitable tasks, or obtain a concession. Supply commitments and continuing development costs can remain when new equipment orders or net prices soften. The first disappointment can be weaker growth or lower cash retained, while existing contracts and creditors continue to be paid.

Evidence is not evenly distributed across those propositions. Existence and technical feasibility of selected alternatives are relatively well supported. Exact realized all-in savings, the private terms governing the two clients and the provider's architecture-specific return are not. The contract calculations are strong as conditional accounting; their scenario prevalence is unknown. Reported cloud profitability is strong favorable evidence at a broad perimeter, not proof that all the new assets earn their cost of capital.

The supported overall change is therefore two-sided but not empty agnosticism: the commercial foundation looks more credible, while automatic preservation of incumbent supplier rents looks less defensible. The investigation rejects the idea that a working alternative necessarily destroys the incumbent and rejects the idea that useful market growth automatically validates every incumbent ownership expectation. It identifies exactly where new measurements could separate rational uncertainty from excessive pricing.

9.3 What observations would change the assessment

A matched post-migration invoice, net of credits and remaining obligations, together with unchanged accepted quality and response performance would establish whether the buyer actually retained the saving. A subsequent voluntary renewal or changed routing allocation would show whether the alternative survives a new purchasing decision rather than an introductory offer. A lower bill without comparable accepted output would weaken the substitution inference; equal service at a durable lower complete bill would strengthen it.

Production capacity at the required tail latency, including standby and peak traffic, would determine whether the alternative remains cheaper near a replica boundary. Better software, batching or right-sizing can move that boundary favorably. More retries, validation burden or paid replicas can move it adversely. Neither direction is assumed from a marketing average.

Finally, architecture-linked receipts and costs through a funded replacement cycle would distinguish buyer gain from provider return. Successful chip-performance milestones, timely capacity delivery and sustained net cash after renewal would strengthen the favorable supplier explanation. Repeated concessions or unproductive commitments without compensating cost gains would weaken it. A whole-company revenue increase alone does not resolve that allocation.

These are discriminating observations, not prerequisites for saying anything useful now and not requests to obtain private information. The present record already establishes real feasible substitution, the conditional cash consequences of published purchasing mechanisms, and a material requirement to invest in supply.

9.4 Coverage and generalization

The principal evidence is a within-AWS inference comparison and a separately documented real-time inference migration. The large Trainium relationship widens the evidence of production scale and investment, not the matched performance sample. The cases do not measure frontier training economics across architectures, every long-context or memory-bound workload, sovereign/on-premises procurement, or the complete economics of AMD and Alphabet alternatives.

The earlier investigations supply customer value, recurring applications, payer funding, hosting, completion, financing, claim allocation and ownership-price contexts. Those are connected economic layers, not independent observations to add into a sector probability. This chapter does not extrapolate a small workload saving to all accelerator sales, a cloud segment margin to a chip margin, or one investment facility to the financial system. It does not sum intermediate transactions or commitments into a “bubble size.”

The question answered is narrower than “which chip wins?” and more useful than a ranking: a buyer can obtain adequate work elsewhere, but the gain is divided by complete service requirements, existing contracts, supply investment and the next negotiated price. That division can disappoint an ownership expectation even in a successful and expanding AI economy.

10. Methods, source status and reusable calculations

10.1 Evidence and decision clocks

The source-inspection date is 16 October 2026. The principal engineering comparison was published 18 December 2024, the production contrast 6 May 2026, and the financial records cover June 30 or July 26, 2026. The AWS price-change event is from June 2025. Current service pages and the terms dated October 13, 2026 explain available contractual mechanisms; they do not establish the clients' earlier or current private agreements. The combination is explicitly a cross-date decision model, not a current head-to-head benchmark or an ex-ante backtest.

Source incentives survive the comparison. Loka sells implementation expertise; AWS sells both competing instance families and promotes its custom silicon; Tomofun supplies a customer engineering account; NVIDIA promotes its integrated platform; Anthropic reports its own supply strategy. Original financial statements support reported accounts, not all causal interpretations attached to them. A promotional source can establish an attributed deployment and still leave full cost, quality and economic return unmeasured.

The two deployment accounts are purposively selected for documentary fit. They are not a representative sample, independent trials of the same intervention or a basis for weighting an expected industry discount. The difference between their reported reductions is retained, not averaged. No paid service experiment, private invoice access, customer contact or new hardware benchmark was performed.

10.2 The complete-cost alternative when private amounts are absent

The calculation uses the best measured core and solves the amount remaining for other necessary work, rather than assigning arbitrary wage, migration or electricity estimates. Unknowns are explicit in the data companion. They include absolute SDXL service times, acceptance fractions, actual migration costs, the clients' contract status, Tomofun's matched old/new total bill and AWS architecture-specific cost. Blank values mean unobserved, not zero.

Zero migration cost in the reference isolates the contractual mechanism; it is not a finding that migration is free. The $5,000 migration example, 10% concession, twelve-month horizon, three-month lock, 12-second GPU time and replica loads are labeled assumptions. The published 17.5% and $14,353 statements are rounded source quantities. Arithmetic precision makes the identities inspectable but cannot improve their empirical precision.

Common service costs cancel only under the stated equality assumption. A cost difference, missing capability or quality change must be introduced on the correct boundary. The model treats repeat attempts as work and capacity; it does not charge them again as an unexplained cash penalty. Purchased host capital is embedded in rental charges; host-side investment is examined separately, without charging the renter that capital a second time.

The financial reconciliation is an accounting identity, not an explanation of the residual. The AWS investment comparison deliberately retains noncash additions and allocation limits. No architecture's lifetime return, current available financing balance or value of scarce upstream capacity is estimated by dividing a broad balance-sheet total by a chip count.

10.3 How to reproduce or change the comparison

substitution-data.csv contains numeric observations, published terms, explicit analytical assumptions and blank unobserved quantities with source, unit, date, entity and locator. substitution-calculations.py uses only Python's standard library. It reads the matching CSV, prints the reconstructed cost account, contract alternatives, concession test, finite-capacity examples and financial reconciliations, and makes no network requests.

python substitution-calculations.py --data substitution-data.csv
python substitution-calculations.py --data substitution-data.csv --self-test
python substitution-calculations.py --tests-only

The script finds the matching CSV beside it, including when both filenames have the same prefix. Results print as JSON for inspection. The script makes no network requests and writes no new file unless its output is redirected.

Changing the relative cost, acceptance, contract term, scope or dated financial input changes the result's meaning. Preserve the corresponding source status and period rather than giving an old number a new observation date. The code refuses an unobserved value when it is requested as a numeric input. Numerical and input checks validate the implementation; they do not verify private payments, benchmark reproducibility, contract enforceability in a dispute or future demand.

Sources, dates and inspection limits

Source inspection boundary: 16 October 2026. The dates below describe the original events, terms and records. Attributed deployment results are not independent replications.

1. Loka, “Productionizing AI Image Generation with Inferentia2,” 18 December 2024, original article. Deployment, benchmark method and the numerical statements in “The Bottom Line.” Accessible body inspected October 16, 2026. Cost/time figure images were unavailable in the cited source inspection; no absolute times or chart values were inferred. Implementer-authored, not independently replicated; the ten-million-image figure is an extrapolation.

2. AWS / Tomofun, “Cost-effective deployment of vision language models for pet behavior detection on AWS Inferentia2,” 6 May 2026, original engineering account. Deployment, migration wrapper, serving architecture, reported saving and concurrency discussion. Body and the published concurrency chart inspected. No raw logs, prior GPU identification, complete invoice ledger or independent acceptance study obtained.

3. AWS, Amazon EC2 G5 instances, configuration table inspected October 16, 2026. Used for the identified G5 bundle, not an observed customer price or an owned-system acquisition cost.

4. AWS, Amazon EC2 Inf2 instances, configuration table inspected October 16, 2026. Bundle capability is distinct from achieved application performance, regional availability and realized billing. No displayed rate without a matched regional/plan definition enters the cost comparison.

5. AWS, Compute Service Level Agreement, page dated May 25, 2022; inspected October 16, 2026. Instance/region availability definitions, topology, exclusions and credit provisions. Not an application-quality or tail-latency guarantee.

6. AWS, Savings Plans FAQ, inspected October 16, 2026. Hourly commitment, duration, family/region versus broader Compute eligibility, payment schedules, return eligibility and absence of a capacity reservation. The actual clients' plans were not obtained.

7. AWS, Service Terms, updated October 13, 2026; inspected October 16. Sections 5.4.1–5.4.2 and relevant capacity provisions. Published legal terms are read with the current FAQ; no private amendment, negotiated exception or legal adjudication is inferred.

8. AWS EC2 User Guide, On-Demand Capacity Reservations, inspected October 16, 2026. Immediate versus future-dated capacity, commitment and modification conditions. A discount mechanism does not itself create delivered capacity.

9. AWS EC2 User Guide, Capacity Reservation pricing and billing, inspected October 16, 2026. Unused versus used reserved capacity and matching-price treatment; no double charge for the same consumed slot is inserted.

10. AWS News Blog, “Announcing up to 45% price reduction for Amazon EC2 NVIDIA GPU-accelerated instances,” 5 June 2025, original announcement and table. On-demand effective June 1; new Savings Plan purchase boundary after June 4. The named P4/P5 families differ from G5; neither wholesale-price causation nor repricing of every old commitment is established.

11. Anthropic, “Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute,” 20 April 2026, updated April 21, original buyer announcement. Existing Trainium2 use is distinguished from prospective capacity, generations and funding. Not a matched Inferentia cost study or a current October capacity census.

12. Amazon.com, Form 10-Q, quarter ended June 30, 2026, original filing. Revenue-recognition/commercial commitments; Note 2, Financial Instruments, Anthropic investments/facility; segment sales, income, PPE, additions and D&A; energy arrangements and MD&A. Financial values retain their statement dates. Net additions include noncash activity; investment ceilings, contractual performance and cash receipts differ.

13. NVIDIA, Form 10-Q, quarter ended July 26, 2026, original filing. Note 10 supply/capacity commitments, inventory provisions, R&D and relevant risk discussion. Inventory-only provisions are not the broader inventory-plus-purchase-obligation measure. The same dated source also supported earlier research; its reuse here is a supply constraint, not a refreshed payment observation.

14. NVIDIA, “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,” 1 October 2026, supplier explanation, inspected October 16. Used to test the economic argument about productivity, continued use, workload breadth and demand. Linked performance/resale analyses were not adopted as matched measurements or independently audited here.

15. Second-cycle recovery, dated investigation with September 29, 2026 evidence boundary and a March 31, 2025 Austin model anchor. Earlier investigation. Its own cost calibration and dates remain separate from this comparison.

16. Buyer demand and budget substitution, dated investigation, October 5, 2026 evidence boundary. Earlier investigation. Used for the distinction between a future purchase, sunk implementation, expanded output and actually avoidable expenditure; its buyer-specific fees and uplift do not calibrate this deployment.

17. Efficiency, competition and full-cost returns, dated investigation, October 6, 2026 evidence boundary. Earlier investigation. Accepted-output, retry, standby and complete-cost conventions; its OCR model and denominator dispute remain their own dated findings.

18. Ownership prices and accumulated economics, dated investigation, October 8, 2026 evidence boundary, October 7 price references and distinct July/June claim inventories. Earlier investigation. Used for the funded supplier scale/margin/renewal combination and favorable alternatives, not as an observed future margin or a new stock-price target.

19. Customer funding and repeat investment demand, dated investigation, October 15, 2026 source-inspection boundary. Earlier investigation. Customer, funding and collection distinctions remain intact. Anonymous NVIDIA customers are not mapped to the deployments here; the unidentified collection adjustment is not set to zero.