Every Sentence Is True: The Real Cost Accounting of an AI Video Course
Sep 24, 2026 · @Mythogen
Ah Zhe is thirty-two. He has spent seven years editing video as a freelancer. Starting last year, the very first question clients asked shifted from "When can you deliver?" to "Can this be made with AI?"
His first encounter with that instructor was through a free online course: a few slide decks accompanied by workflows, prompt templates, and sample videos, with every single step meticulously laid out. He followed along and successfully generated his first talking-head video. In the comments section below the post, he left a note of gratitude alongside hundreds of others: So generous, the instructor is an angel, finally got it to work. In that very same thread, some were asking when a paid class would begin; others were already clamoring to sign up.
A few days later, the instructor published an analytical post doing some math. A subscription to a large language model costs about $100 a month and lasts the entire month; cloud-based video generation costs the same, but if you reroll aggressively, your quota evaporates in three days. For individuals, rather than running language models locally, it makes far more sense to run open-source video models on a local machine—a mid-to-high-end GPU can brute-force the workload. Converted against cloud pricing based on his casual daily generation volume, running it locally saved over $500 a month.
Ah Zhe did the calculation: saving $500 a month meant breaking even on hardware in just over half a year.
He bought a pre-built desktop rig for over $3,000, equipped with a 16 GB VRAM graphics card—the post had noted that this card "could also pull it off." Then he signed up for the instructor's paid masterclass: a few hours long, priced at $400. The class was solid. How to connect nodes, how to adjust parameters, which folder to store model checkpoints in—everything was explained with crystal clarity.
Over the month following the course, he practiced diligently. The footage moved, but never with the silky fluidity seen in the instructor's portfolio; whenever foliage entered the frame, the leaves turned into a smudged blur, and on-screen typography was consistently fractured. He shared his attempts in the student community group. The instructor replied: Reroll more times, practice more.
By the third month, a brand-new open-source model was released. It was vastly superior to the previous one, but 16 GB of VRAM could not run it. In that same month, a free utility went viral in the developer community, enabling AI agents to automate almost all the setup, environment configuration, and debugging taught in the class. One week later, the instructor launched a brand-new course.
Ah Zhe blamed nobody. In the group chat, he said he simply lacked talent and needed to work harder.
In this entire story, not a single person ever uttered a lie.
1. What the Calculation Got Right
To be completely fair: the core logic of that cost analysis was sound.
Subscriptions to language models are fixed monthly fees heavily subsidized by venture capital; whether you use them heavily or lightly, the price tag remains the same. Video generation is fundamentally different: it is a pay-per-pull lottery. For a single prompt, one often has to reroll dozens of times just to extract a few usable seconds of footage. Cloud platforms meter every single pull; generating locally reduces the marginal cost of each additional pull down to electricity. The more you reroll, and the more lenient your tolerance for visual imperfections, the more economical local compute appears.
This judgment was not nonsense. For a certain profile of user, it is entirely accurate. The problem is that it tallied only half of the balance sheet.
2. The Half It Left Out
The first expense missing from the ledger is hardware depreciation. The post contrasted a monthly recurring fee against a one-off graphics card purchase, but purchasing hardware is a capital expenditure that must be amortized over its usable operational life. In the realm of local video generation, that lifespan is brutally short. In times of supply shortages, the market resale price of a GPU may not drop quickly, but its functional obsolescence arrives at breakneck speed. A graphics card that has to "struggle to pull it off" today is already operating right up against its VRAM ceiling; the moment the next generation of models demands one tier higher, it instantly demotes from "usable" to "restricted to running legacy models." If a $3,000 workstation becomes functionally obsolete in eighteen months, its monthly amortization already equals or exceeds the cloud subscription it was intended to replace.
The second missing expense is workflow depreciation. With every major model update, quantization versions, custom nodes, drivers, and parameters often have to be re-tuned from scratch. This debugging time represents a very real cost—it simply does not appear in currency form, and therefore never makes its way into any simplified financial formula.
The third is even more insidious: the supposed "$500 monthly savings" is largely phantom. The instructor's calculation multiplied his casual, habitual "generate whenever an idea strikes" volume by commercial cloud unit rates. But that volume only exists precisely because pulling locally incurs no per-token fee. If every generation were metered and billed, no rational creator would pull at that frequency. Multiplying zero-cost consumption volume by commercial pay-as-you-go pricing treats non-existent expenditures as realized income.
There is another form of depreciation that fewer people notice: advice depreciates. Earlier, the same author had recommended a dedicated mini PC designed for local AI to run local LLMs. Now, his cost analysis pointed out that local LLMs struggle to match cloud efficiency—that even that recommended mini PC labored under the strain—and advised individuals to pivot to video models instead. Pivoting advice carries zero cost for the person giving it; the pivot itself even serves as fresh content. But the reader who bought hardware based on the preceding recommendation is permanently locked into yesterday's advice.
Cloud pricing is also far from static. The instructor's post itself noted that whenever a major open-source model drops, cloud providers immediately launch price cuts and promotional credits. Renting cloud capacity preserves the optionality to re-evaluate your stack every month; buying hardware exercises that option prematurely, betting that cloud prices will not fall faster than hardware depreciates. In periods of extreme technological volatility, flexibility is the most valuable asset, and a capital buyout is the abandonment of flexibility.
Finally, the exact same machine yields completely different balance sheets depending on whose hands it sits in. For someone making a living through AI instruction and consulting, a single machine simultaneously produces course curriculum, social media posts, hardware benchmark reviews, and client deliverables. For them, equipment is productive capital, and upgrading machines is itself raw material for content creation. For Ah Zhe, the revenue column remains zero until he actually sells a client project. The ledger was entirely sound from the instructor's vantage point; the error lay in handing that ledger unaltered to someone operating under an entirely different revenue structure.
3. The Other Ledger in the Comments
What made that post truly worth studying was its comment section.
An industry peer specializing in commercial advertising left a comment stating that commercial projects overwhelmingly demand native 4K. Upscaled footage routinely fails client quality control, with smudged foliage and fractured typography being the most frequent grounds for rejection. While cloud rendering burns cash, corporate clients reimburse the budget, so agencies simply burn through it. The instructor's reply was remarkably candid: on genuine large-scale projects, financial backers provide substantial budgets for you to push limits; however, before landing such contracts, ordinary creators must first build a portfolio. In another reply, he mentioned that on-screen typography typically requires leaving blank composition space to be composited during post-production.
This was not deceit. A comment section is an open dialogue, and the author answered straightforwardly. But it illustrated one crucial truth: the conclusion of the post depends entirely on how far the reader scrolls. Those who read only the main text conclude that "local compute is cost-effective"; only those who dig through the replies discover that commercial client delivery takes an entirely different road.
The profitable ledger relies on the cloud, treated as operating expenses reimbursed by clients. The hobbyist ledger recommended to individual learners relies on local compute, where capital outlay, depreciation, and debugging time are swallowed entirely by the student. The former can pass costs downstream; the latter absorbs sunk costs.
And such a comment section is already a rare exception. The post Ah Zhe read was worth studying precisely because the author permitted such exchanges and was willing to engage. In the vast majority of commercial creator communities, comments questioning unit economics or demanding production specifics do not survive: moderators delete them, ban the accounts, or remove threads citing "negativity" or "unfriendly vibes." My own analytical essays have been simultaneously censored in several major AI communities; as documented in my previous analyses, the engagement metrics of certain large communities are artificially sustained by employee burner accounts and affiliate networks. What survives the filter is gratitude, enrollment inquiries, and praise.
Consequently, the other ledger is not simply buried in the comments—it is denied existence altogether. Information asymmetry does not merely stem from what sellers omit; it stems from community curation that prevents buyers from asking in the first place. The public discourse a student evaluates before parting with their money has already been heavily sanitized.
4. The Five Things the Classroom Cannot Teach
Return to Ah Zhe. That he failed to replicate the instructor's visual polish was not due to a lack of diligence. The instructor's portfolio pieces are the cumulative product of five distinct layers, and a course spanning a few hours can only ever teach the first.
- Tool Operation: Node chaining, workflow orchestration, and parameter adjustments. This was the scope of the class, and it was taught well.
- Cloud Compute: As the author acknowledged in his post, several flagship clips in his portfolio were rendered on cloud clusters, completely maxing out his enterprise quotas. The class, however, taught local execution.
- Reroll Volume: The author acknowledged that massive quantities of casual generations are never published. A portfolio is the residue of aggressive curation; the hundreds of warped, failed generations behind it appear neither in the showcase nor in the lesson.
- Post-Production: Composited typography, localized inpainting, and color grading were executed in dedicated editing suites after the AI generation step, well outside the scope of the node tutorial.
- Judgment: Recognizing which roll possesses usable structural bones, and precisely how to adjust a prompt to correct trajectory, is tacit knowledge accumulated through thousands of iterations. It cannot be transferred in a single afternoon.
The production conditions of the showcase piece lie fundamentally outside the boundaries of the curriculum. The student beholds the sum of all five layers, but purchases only the first.
5. Harder to Expose Than a Scam
Courses that promise to make you a millionaire are relatively easy to dismantle: they promise a concrete outcome, and when that outcome fails to materialize, the claim is falsified.
The class Ah Zhe attended was different: it never promised anything. The showcase was real, the tools were real, the curriculum was real, and the cost comparisons in the post were largely accurate. The disparity never manifested as a stated sentence; it hid in the gulf between showcase and delivery. The recruitment pitch exhibited the finished work; the classroom delivered the bare tool. The cloud quotas, reroll volume, post-production, and tacit judgment that linked the two were never itemized on either side.
This represents a more sophisticated paradigm than an outright scam: it requires no lies, only orchestration. Every single sentence is true, yet together they synthesize a false expectation. Consumer protection regulations cannot reach it, because not a single statement is legally actionable. Critics struggle to articulate objections, because every factual premise advanced by the instructor stands up to scrutiny.
And precisely because it requires no lying, it proliferates in numbers far exceeding traditional scams. An outright pyramid scheme requires an organization, a brand, and a deceptive script, making it a centralized target that can be exposed. This paradigm requires only an individual "creator with a portfolio"—and generative AI has compressed the threshold for possessing a portfolio lower than ever before. The result is an army of hundreds of fragmented individual operators, leaving no central headquarters to challenge.
6. Price Was Never the Issue
Learning is essential; the objective is to avoid learning blindly. And when evaluating an educational offering, price has never been a reliable diagnostic.
Consider scale model painting and modification masterclasses. Led by master craftsmen, some cost twenty dollars for a technique demo, while others cost thousands for multi-month mentorships. Their teaching styles diverge wildly, as does their quality. Yet they share three indispensable attributes: transparent pricing, predictability, and trackability.
Pricing is transparent because the requisite materials, tools, and expected time commitments are documented upfront; unexpected structural costs do not suddenly surface after enrollment. It is predictable because the master's showcase was executed using the exact materials and techniques demonstrated in class; what the student sees is identical to what the curriculum encompasses. It is trackable because when a student fails to achieve the master's level of execution, the bottleneck is self-evident: manual dexterity and accumulated hours on the brush. That gap is honest, visible, and quantifiable. The student's tenth build will be visibly better than their first.
AI courses become problematic not because their tuition is high, but because all three mechanisms fail simultaneously. Total cost is concealed not merely in GPUs, cloud top-ups, and hardware depreciation, but in the pre-configured environments that worked seamlessly on the instructor's setup but break upon install at home; in the specialized nodes, proprietary checkpoints, and workflow packs upsold after class; in subscriptions to upscalers and editing suites; in commercial licensing ambiguities for open-source weights; and in the recursive rerolling triggered every single time a real-world client requests a revision. The showcase relies on cloud power, immense reroll volumes, and post-production outside the curriculum, rendering outcomes completely unpredictable. And when a student cannot replicate the results, the deficiency is not manual craftsmanship, but an array of invisible capital and compute inputs they cannot see. Unable to track the gap, they have no choice but to blame themselves.
In manual craftsmanship, output roughly equals skill multiplied by time; the showcase piece is inextricably bound to instructional content. In generative AI, output is a lottery: a showcase can be constructed by pouring hidden inputs into the machine, completely decoupled from the bare mechanics taught in class. This explains why such courses have exploded across social media: technology has, for the first time, enabled the systematic, large-scale decoupling of what can be displayed from what can be taught.
Yet technology does not mandate this decoupling. As long as an instructor provides raw, unedited outputs, fully reproducible seed prompts, and dual workflows for both local and cloud setups, this gap can be drastically narrowed. Indeed, some educators already do this. Technology permits opacity, but it does not demand it. A course that fails to provide transparency does not fail out of impossibility; it chooses opacity by design.
7. When the Genuine Article Backs the Trend Course
In The Efficiency Trap (Postscript), I mapped course creators along a spectrum: on one end lay genuine skill at fair pricing; on the other lay packaged anxiety arbitrage. At the time, I cited a course I had personally taken and found genuinely useful as an exemplar of the healthy end of the spectrum.
That classification requires revision. The error was not in acknowledging that the specific class was valuable, but in extrapolating that "this class is useful" meant "this instructor belongs entirely to the healthy end of the spectrum." In reality, the exact same instructor can simultaneously offer rigorous, high-utility masterclasses alongside opportunistic, trend-chasing crash courses. The spectrum does not run between separate individuals; it runs directly down the product catalog of the same creator.
This hybridity is itself the primary engine of concealment. The rigorous course acts as a credit guarantor: students take it, find genuine utility, and build trust. That trust is subsequently ported over to the next offering. The genuine article endorses the trend-chasing course—and the endorsement is forged by the student's own authentic experience, making it far more persuasive than any paid advertisement.
The instructor Ah Zhe encountered is the clearest illustration of this dynamic. His free introductory course was published openly on his website, complete with dual local and cloud workflows, copy-paste prompts, and actual generated video files. The page explicitly noted: This video is the unedited output generated by this workflow; following these steps should yield a comparable result. The materials were unified, allowing anyone to verify them page by page. He even openly critiqued industry peers whose tutorials tied students to a single proprietary cloud platform, leaving local installations broken. In those respects, he operated far above the industry standard.
His paid course, however, followed an entirely different architecture: customized slide decks tailored to private corporate workshops, with no public syllabus, no transparent pricing, and no reproducible source files. The former was structurally verifiable; the latter was structurally unverifiable. And delivering the fidelity of his public portfolio within a few hours of classroom instruction was, as broken down earlier, physically impossible.
This instructor clearly understood what transparent pedagogy looked like, having demonstrated it with his own hands. The problem was not a lack of comprehension, but that the standard was not applied to the revenue-generating side of the business. The trust established by the free course was anchored in standards he had set himself; the paid course enjoyed that trust without being subjected to the same scrutiny. Evaluating that paid offering requires no external moral yardstick—it requires asking only one question: Did it adhere to the standards he promised in his free class?
The genuine article shields the paid offering across at least five critical dimensions:
- The Timeline Gap: Deep curriculum may reflect years of practical industry mastery, whereas hot new topics may have emerged only months prior, meaning the instructor is merely one step ahead of the students.
- The Verifiability Gap: Free tutorials can be audited line by line, whereas paid deliverables exist only as untouchable showcases.
- Content Lineage: Much of the paid material simply repackages official documentation, open-source repositories, and freely available community posts.
- Survivor Bias: Testimonials originate exclusively from satisfied outliers; those who fail to produce results remain silent, overwhelmingly assuming personal inadequacy.
- Shelf-Life Disparity: Different courses depreciate at radically divergent speeds, yet housed under the same personal brand, the student cannot discern the decay.
Why instructors adopt this model invites multiple explanations. It may be deliberate funnel optimization: utilizing free content as a trust on-ramp and monetizing through paid cohorts. It may be distribution pressure: recommendation algorithms ignore legacy topics, and failing to chase trending keywords means vanishing from feeds. It may be genuine cognitive bias: masters in one domain easily convince themselves they are equipped to teach adjacent territories. It may simply be economic survival: a single curriculum saturates quickly, and teaching for a living demands an endless pipeline of novel products. Or it may merely be market responsiveness: students themselves actively demand the class.
Every explanation is plausible, and several can operate simultaneously within the same individual. This is precisely what makes the model more elusive than fraud: intent always possesses a rational justification, rendering accountability impossible to assign. Yet whether the instructor acts out of calculation, systemic coercion, or sincere conviction, the outcome on the student's side remains identical: trust is extracted, costs are absorbed by the learner, and failure is internalized as personal deficiency. Intent can have ten explanations; consequence has only one.
8. Institutions Buy Exactly What It Sells
Were this merely an isolated market quirk, it would not operate at such scale. Its prevalence is sustained by two formidable institutional pillars.
The first is government. In my previous writing, I analyzed how public subsidies act as macro-amplifiers for this ecosystem. We must take that analysis one step further: government workforce development programs evaluate instructional hours delivered, graduate headcount, and certificates issued—never whether a graduate can deliver production-grade work. Teaching a software tool in a weekend cohort produces precisely the requisite hours and completion metrics. What the institutional procurement apparatus buys aligns seamlessly with what these courses sell. Under this framework, that Ah Zhe cannot reproduce the instructor's portfolio will never surface on any institutional KPI dashboard; that he received a completion certificate will.
The second pillar is enterprise. The narrative that "local compute is cost-effective" directly benefits GPU and desktop workstation sales. The cost analysis Ah Zhe read advised readers to purchase mid-to-high-end GPUs; the author maintained an open collaborative relationship with hardware manufacturers, which was not disclosed in the post. We do not know the commercial nature of that collaboration, nor is there reason to assume the post was sponsored content; yet readers evaluating hardware economics have a fundamental right to know that context. More broadly, such narratives require no formal sponsorship at all: they constitute an organic, uncompensated sales funnel for hardware manufacturers.
Another structural pillar is the transfer of institutional legitimacy: government-partnered workshops, corporate training rosters, adjunct university appointments, and vendor keynote titles. Each layer of institutional endorsement is ported onto the private course, transforming a profile of "self-taught creator with a few workshops" into an authoritative credential. The student sees an expert vetted by institutions; what they do not see is that those institutional validations never measured production delivery competence.
9. Involution Engineered by the Course
If one studies purely for personal enjoyment, this section can be bypassed: market saturation does not diminish the joy of artistic expression. But Ah Zhe did not learn for a hobby; he learned to freelance. And the marketing copy for these courses specifically targets freelancing, monetization, and secondary income.
Preceding sections examined costs borne by the student directly. There is another ledger far more difficult to perceive, because it resides not on the student, but inside the market they are about to enter.
The more successfully a tool course teaches, the more people master the mechanical workflow; the more people who master it, the cheaper that specific skill becomes in the market. Each graduating cohort becomes the direct competitor of the previous cohort. The only actor standing outside this zero-sum loop is the course creator: his revenue derives from enrollment volume, not from competing inside the saturated labor market. Ah Zhe's instructor explicitly notes on his website that his primary occupation is AI educator. For him, more students is strictly positive; for the student, more peers is strictly negative. The identical enrollment metric carries diametrically opposite economic implications for the two parties.
Tiering within the client pool compounds this friction. High-budget enterprise contracts flow to established studios with existing client networks, track records, and production pipelines. Commercial delivery standards—native 4K, cross-shot character consistency, vector-clean typography, and clear copyright indemnification—stand vastly higher than the baseline for personal portfolios. Newcomers can access only the bottom tier of low-paying gigs, and that bottom tier is precisely where each wave of course graduates converges. The course sells an admission ticket, but the arena floor behind the turnstile has already been overcrowded by earlier cohorts.
A parallel involution is driven by technology itself. Visual effects that require complex node chains and parameter tuning today may be consolidated into a single button in the next model update; environmental setup and package conflicts that demand hours of instruction today will be automated by AI agents tomorrow.
While writing this essay, I reviewed a newly open-sourced workflow migration utility. Its purpose is to deploy AI agents to ingest ComfyUI workflows built on cloud platforms and re-architect them to run locally. The agent audits missing nodes and checkpoints, searches community repositories, assesses security risks, converts dependencies, validates execution, and outputs a detailed ledger of substitutions. The guide is completely free and licensed under an open-source framework.
This technical guide is well worth reading, because it inadvertently codified the hidden costs discussed throughout this essay into rigorous technical documentation. When cloud workflows are ported locally, they almost never execute out of the box. Proprietary platform nodes cannot be migrated. Assets uploaded by the original creator appear as cryptographic hashes that cannot be retrieved, forcing local users to source approximations. For every substituted component, the audit report must explicitly flag that the output no longer mirrors the original creator's result. It cited a real-world case: an advanced workflow chaining fifteen LoRAs required twenty minutes of automated dependency resolution and conversions, followed by downloading a base model approaching 20 GB. It cautioned that updating a single pre-installed Python library can, without throwing an error message, subtly alter the output of existing workflows.
These are operational realities that industry practitioners grasp intuitively, but which laymen have no means of knowing.
The utility simultaneously demonstrates the trajectory of technical involution. The solitary layer that tool courses successfully deliver—nodes, models, environments, and debugging—is being handed off to AI agents via a free Markdown document. The core curriculum of a multi-hundred-dollar masterclass has been superseded by an automated file, authored by an industry educator himself.
This is not an accusation. Open-sourcing such a utility is genuinely generous, and its ethical standards are rigorously transparent: if an asset cannot be located, it reports that both search paths failed, rather than claiming the file does not exist; if output fidelity cannot be evaluated, it defers judgment to the human user, rather than substituting an arbitrary benchmark. What it demands of AI is precisely what this essay demands of educational courses: articulate what is missing, document what was substituted, and state what cannot be guaranteed.
The tragic irony is that this transparency appears in instructions written for machines, yet remains absent from sales pages written for human students.
Market involution depresses what students can earn after graduation; technical involution accelerates the obsolescence of what they learned. Compounded together, students confront double depreciation: their skills grow cheaper by the month, while their knowledge grows shorter-lived by the day. And both forms of depreciation serve precisely to manufacture demand for the next course.
10. Courses Depreciate, and Depreciation Is the Business Model
Graphics cards depreciate, workflows depreciate, advice depreciates, and mechanical operations learned over a weekend depreciate just as quickly. When the next foundation model drops, nodes shift, parameters reset, and classroom steps no longer apply. To the student, this represents capital loss; to the course seller, it represents repeat business.
Outright get-rich-quick scams must continuously prospect for fresh marks, because those burned once do not return. This paradigm requires no such churn. When a new model arrives, the existing cohort of trusting students assumes that they are the ones falling behind. The shorter the shelf life of a curriculum, the higher the repurchase velocity. Students spend capital racing against the very velocity that compels them to spend capital again.
In his third month, Ah Zhe observed three simultaneous phenomena: a new foundation model, a free tool that automated what he had paid to learn, and a brand-new masterclass. His GPU could not run the first, the utility rendered his tuition redundant, and his lingering trust made him contemplate enrolling in the third.
Conclusion
This essay names no individuals. The omission is not out of caution, but because naming names reduces a systemic issue to personal character, when the dynamic has never depended on the virtue of any single actor. Ah Zhe's instructor accomplished much that deserves genuine praise: a free introductory course that can be audited step by step, and an open-source technical guide whose transparency is uncompromising. The problem resides in the structure: a machinery that requires no one to utter a lie to reliably siphon capital from one side of the table to the other.
This essay is not an argument against learning. Learning is essential, and paying for quality instruction is entirely justified; how much one pays is a subjective valuation. What must be avoided is learning blindly.
Nor does learning need to justify itself through income. Painting, playing the cello, or studying chess are inherently worthwhile pursuits; no one asks how much a recreational piano student earns from their scales. The friction emerges exclusively in courses that teach the depth of a recreational hobby, yet market the expectation of professional income.
Rather than asking "Is this instructor trustworthy?", one should ask five diagnostic questions before enrolling in any commercial course:
- Is the total cost transparent? Are hardware demands, cloud allowances, post-production requirements, third-party subscriptions, and subsequent upgrade costs disclosed before payment?
- Is the outcome predictable? Can the visual fidelity showcased in promotional materials be produced strictly within the boundaries of what the class teaches?
- Is the performance gap trackable? When your output falls short, can the deficit be traced to your personal craftsmanship, or does it stem from hidden capital inputs?
- Where does the instructor's primary income originate? Does it derive from delivering client work in the market, or from selling courses to students?
- Does the course sell a hobby or an income? If it implies freelancing, commercial viability, or second-career monetization, have recent graduates demonstrably achieved this? If it carries no such implication, return to the first three questions: Is what you learn genuinely what the promotional copy promised?
A course that answers all five satisfactorily can be expensive and still be exceptional. A course that cannot answer them is overpriced even if it is cheap.
There is an even simpler litmus test: copy these five questions and post them in that instructor's public community. If the comment remains up and receives thoughtful engagement, that is your answer. If the comment is quietly removed, that too is an answer.
Ah Zhe told his study group that he simply had not worked hard enough. He was mistaken, but his mistake was not a lack of effort. He believed he was purchasing the cumulative product of all five layers; he purchased only the first. The remaining four, he is now paying for with his own time, his graphics card, and his self-confidence.
Related Reading: The Efficiency Trap, The Efficiency Trap (Postscript): You Thought You Won, Capital Game series