One Terawatt, No Timestamp: What TeraFab’s Missing Units Say About the AI Compute Endgame

Silence in the slasher was the first warning sign. That line has anchored every protocol audit I have ever filed. It is the line I think of when I read an infrastructure announcement that is technically loud but structurally quiet. A missing field. A dangling unit. A claim without a date. So when the TeraFab material crossed my desk — a fragmented report claiming that a new Musk-affiliated compute initiative would deliver one terawatt of AI capacity, with 75 percent going to AI spacecraft and 25 percent going to Optimus — I did not start with the terawatt. I started with the silence. There was no publication year. There was no primary link. There was no author attribution. There was no direct quote. There was, instead, a number so large that it displaced all other questions: 1 TW. In my 26 years of reading infrastructure claims, the largest numbers are rarely the most informative. They are usually the safest place to hide the missing details. The proof is in the unverified edge cases. A terawatt is not a compute target. It is a civilization-scale electrical claim. The global data center fleet currently consumes roughly 460 to 500 terawatt-hours per year, representing an average sustained power draw of about 52 to 57 gigawatts. Those numbers come from multiple industry trackers, and they already include hyperscaler campuses, colocation facilities, enterprise server rooms, and a long tail of crypto mining warehouses. If TeraFab were to be built as literally one terawatt of installed power capacity, it would add nearly twenty times the current average load of the entire global data center industry onto a single infrastructure project. No date. No grid interconnection queue. No power purchase agreement. No environmental review. No corporate entity. No audited balance sheet. Just the number. I have spent years auditing systems where a single omitted variable turns a sound design into a catastrophic trust failure. The Ronin bridge did not fail because of a random bug; it was engineered to trust a small set of validator keys, and the proof was in the unverified nonce reuse that everyone missed. TeraFab is being sold as a breakthrough in physical AI infrastructure, but the early documentation carries the same pattern: a massive concentration of resources, an interlocking set of corporate entities, and a narrative that asks outsiders to accept the architecture on faith. The first warning sign is not the scale. The first warning sign is the absence of a timestamp. Let me be blunt about what we actually know. TeraFab, according to the report, is an AI compute project associated with Elon Musk’s network of companies. It reportedly plans to allocate approximately 75 percent of its compute to AI spacecraft and 25 percent to Tesla Optimus. The report itself rates its confidence at C. It explicitly admits that the year is missing, the primary source link is missing, the author identity is missing, and any direct quotation is missing. The report is an interpretation built on industry common sense and extrapolation. That does not make it worthless. It makes it a hypothesis in need of verification. I have learned to take such hypotheses seriously. In 2017, while the ICO market was roaring and nobody wanted to read a proof-of-stake slashing spec, I spent six weeks manually auditing the Ethereum 2.0 Phase 0 slasher logic and found three state-reversion vulnerabilities. The pattern was always the same: the system looked elegant at the level of happy-path execution, and the dangers lived in the conditions everyone assumed would never happen. In 2020, I built Python simulations to deconstruct Curve Finance’s StableSwap invariant, and the non-linear fee adjustments created hidden arbitrage channels that were invisible to the casual reader. In 2022, I traced the Ronin exploit transaction by transaction through four layers of smart contract interactions and showed that the flaw was not in consensus but in off-chain validator signature verification. In 2024, I ran a custom stress test against Solana’s TPU path and watched finality latency degrade in ways that contradicted every official scaling claim. I have seen what happens when engineers confuse narrative certainty with mathematical proof. TeraFab is not a smart contract. It is not a blockchain protocol. But it is the same species of problem. It asks us to evaluate a promise about enormous compute, physical-world deployment, and interlocking corporate control. The same forensic instincts apply. I want to know which invariants are being claimed, which edge cases are unverified, and which actor is being asked to trust another. A unit error in a press release is a small thing. A unit error in a funding narrative is a signal. And a unit error in an infrastructure claim of this magnitude is the difference between a real data center project and a political symbol. Let us begin with the unit forensics. The report says TeraFab is aiming for one terawatt of compute capacity. The phrase 1TW is ambiguous between installed power capacity and annual energy generation. One terawatt of installed power capacity, if operated continuously for a year, produces 8,760 terawatt-hours of energy. The entire world currently generates about 30,000 terawatt-hours of electricity from all sources combined. Another way to say this is that TeraFab, under the aggressive reading, would consume roughly 29 percent of total global electricity production. Under the conservative reading, the report might have intended one terawatt-hour per year, which would imply an average power draw of roughly 114 megawatts. That is not a speculative aerospace play. That is a plausibly large but entirely conventional data center load. The gap between 114 megawatts and 1,000 gigawatts is not a rounding error. It is the difference between a building and a new energy civilisation. When an infrastructure report carries this level of unit ambiguity, I immediately suspect that the ambiguity is functional. A claim of 1TW generates awe, media coverage, and capital flow. A claim of 1TWh/year generates a meeting with a regional utility. The source material contains no clarification, and that absence is the first technical finding. There is a deeper mathematical problem. If TeraFab were to be constructed as a single 1TW installation, it would need to be located somewhere that can supply primary power with essentially zero interruption. Texas is the obvious candidate because of the existing Musk infrastructure ecosystem and the state’s historically permissive approach to large industrial loads. But the Electric Reliability Council of Texas, commonly known as ERCOT, had a system peak load around 80 gigawatts in recent years. One terawatt of new load represents more than twelve entire ERCOTs. No utility board on earth will sign that interconnection study. No equipment manufacturer is building transformers at that scale. No transmission corridor in North America can carry that current without a fundamental rethink of the grid. The material side of the claim collapses under arithmetic that any first-year electrical engineer can perform. The 75 percent allocation to AI spacecraft is even stranger. If we take the literal 1TW reading, that means 750 gigawatts of compute for satellite autonomy, spacecraft control, and orbital inference. The total mass of every satellite ever launched is under roughly 10,000 tonnes. Fitting 750 gigawatts of AI compute onto orbital hardware would require launching the equivalent of thousands of fully equipped terrestrial data centers into low Earth orbit. That is not an engineering roadmap. It is a statement of intent that floats above physics. If we take the conservative 1TWh/year reading, 75 percent of 114 megawatts is about 85 megawatts, which could power a constellation of satellites equipped with modern inference accelerators. That is still unusually large for the space industry, but it no longer violates the laws of thermodynamics. The reality is probably neither extreme. TeraFab is a name that will be attached to a family of projects: a Texas-based data center campus, a purchase of GPUs from Nvidia or maybe a competitor like Cerebras, a set of power contracts with gas plants or potentially a nuclear source, and a long-term claim on future computing capacity. The 1TW number is best understood not as a current capacity target but as a narrative anchor. It is the number that makes the project unmissable. It is also the number that makes the project impossible to verify. The report’s own confidence rating of C is an admission that the source material cannot sustain the weight of its headline. I have seen this trick before in the blockchain world. Projects announce a theoretical maximum throughput of millions of transactions per second, show a chart with a logarithmic y-axis, and never mention that the latency under load destroys the use case. The proof is in the unverified edge cases. With TeraFab, the edge case is the unit. With TeraFab, the edge case is the date. With TeraFab, the edge case is the legal entity. Who pays for the power? Which shareholders own the asset? Which government authority has jurisdiction over an AI spacecraft with autonomous decision-making? None of those questions appear in the forty bullet points of the report, and that is not an accident. The source report spends considerable effort on the technological route, commercialisation, industrial impact, competitive landscape, ethics, investment, and infrastructure. Only three dimensions are rated high relevance: industrial impact, competition, and infrastructure. The ethics dimension is rated low. That is conspicuous. A project that plans to put AI decision-making on spacecraft, with no human operator loop, raises obvious ethical and legal questions about the use of force, orbital debris avoidance, and accountability. To rate ethics as low relevance is to make an implicit political choice. It says that the scale of the engineering promise is more important than the governance of its deployment. I have audited enough systems to know that governance is not a soft field. Governance is the grid of the protocol. If the governance breaks, the computation breaks. The competition analysis in the source report is useful, but it is also incomplete. It notes that Microsoft, Google, Meta, and Anthropic are building GPU clusters at the scale of hundreds of thousands of accelerators. It is true that TeraFab’s differentiation would not be in general-purpose model training but in the vertical integration of compute with physical-world platforms: Tesla for robotics, SpaceX for launch and space access, Starlink for high-throughput orbital data, xAI for foundation models, and X for distribution. That stack is remarkable. No other company on earth combines a leading EV manufacturer, the dominant launch provider, the largest satellite constellation, a frontier AI lab of arguable quality, and a global social media platform with 500 million monthly active users. OpenAI has Microsoft. Google has DeepMind. Meta has social graph data and open-source weight distribution. Anthropic has safety narrative and enterprise sales. None of them have a rocket factory. That vertical integration is the core of the TeraFab thesis. It is also the core of the threat narrative. The report frames TeraFab not as a proposal but as a weapon in a second wave of competition. The first wave was about GPU count. The second wave is about power capacity and physical-world deployment. Microsoft, Google, and Meta are spending tens of billions of dollars per year on capital expenditures, largely for cloud AI workloads. They are building data centers in places like Iowa, Virginia, Ohio, and Singapore. They are signing power purchase agreements with nuclear operators and investing in small modular reactors. But they are not building spacecraft. They are not building humanoid robots. They cannot put a model into orbit and have it collect its own data. TeraFab, if it exists as described, would be the first attempt to fuse frontier model training with physical-world data collection at the operational layer of the network. There is a specific technical reason why AI spacecraft should matter to someone who thinks on the protocol level. Every satellite today is controlled by deterministic command sequences. A Starlink satellite may have a limited amount of onboard autonomy for collision avoidance, but the intelligence of the network resides on the ground. The ground stations are functionally centralized compute centers with high-latency links. If a spacecraft needs to make a decision about an orbital collision, it cannot wait for a round trip to a server in Texas. The round trip latency to low Earth orbit is tens of milliseconds, which is fast enough for telemetry but not for high-frequency autonomous control. A truly autonomous AI spacecraft must run inference onboard. That requires a radiation-hardened accelerator, a power-efficient chip, and a training regime that can be updated continuously from ground-based models. The 75 percent allocation to AI spacecraft is a signal that Musk is trying to make Starlink the compute substrate of space. Starlink satellites would stop being dumb routers and become an edge inference layer. That is what the source report hints at when it calls this a potential upgrade from communication network to space edge computing network. I find that possibility genuinely fascinating. It is also terrifying from an infrastructure risk perspective. A satellite constellation with onboard GPU inference is a distributed system with a bounded number of nodes, long distance between nodes, and no opportunity for physical maintenance. The failure domain of a traditional data center is contained within a building. The failure domain of an orbital compute network is the entire sky. If a radiation event flips bits across a large portion of the constellation, you cannot simply reboot a rack. You have to coordinate a fleet under resource constraints. The mathematics of consensus in such a system are closer to proof-of-stake slashing than to cloud computing. The report does not engage with this at all. It treats AI spacecraft as a market opportunity and ignores the protocol-level fragility. Let me turn to the Optimus allocation. The source report says that 25 percent of TeraFab compute will go to humanoid robot development. If we assume the conservative 1TWh/year reading and ignore the broader headline, 25 percent of 114 megawatts is roughly 28.5 megawatts. That is enough to power a few thousand H100-class accelerators. A few thousand GPUs is a serious training cluster but not unprecedented. Figure, 1X, and other humanoid robotics companies operate at a much smaller scale, often hundreds of GPUs. A dedicated internal cluster of several thousand accelerators for Optimus would give Tesla a meaningful advantage in data-hungry imitation learning and world-model training. It would not create a durable moat overnight, but it would shorten iteration cycles. On the other hand, if the 1TW headline is taken seriously, 25 percent would be 250 gigawatts, which is so absurd that it can only be interpreted as a branding statement. The unit ambiguity again collapses any attempt to extract an operational forecast from the article. The Optimus advantage is not just compute. Tesla has a fleet of vehicles on the road collecting visual and control data in real-world environments. That fleet is a data generation engine for training autonomous behaviour. The same neural architecture that learns how to drive a car in San Francisco can be repurposed, with substantial modifications, to explore how a humanoid arm should grip a tool. The report correctly identifies this as a transferable engineering asset. OpenAI has no fleet. Anthropic has no robots. Google has a robotics research division, but it does not have a mass-produced electric vehicle base with extensive real-world sensor coverage. The Tesla advantage in embodied intelligence is real, and it is independent of the 1TW fiction. This is the part of TeraFab that deserves close technical attention even if the terawatt is never built. The most interesting governance question, however, is not about Optimus. It is about TeraFab’s corporate structure. The source report says nothing about the legal entity. It suggests that TeraFab is likely a shared compute platform among Tesla, SpaceX, xAI, and X, rather than a wholly independent company. If that is true, then 75 percent to AI spacecraft means Tesla shareholders would be providing capital, directly or indirectly, to a compute facility whose main beneficiary is SpaceX. There is no established market price for AI compute transfer between a car manufacturer and a space logistics company. This is a transfer-pricing nightmare. It is also a shareholder-governance nightmare. In the crypto world, we would call this a conflict of interest manifested in smart contract privileges. In the corporate world, it is simply an unenforceable promise. The source report notes that xAI’s Colossus cluster may remain independent of TeraFab. If Grok is not a consumer of TeraFab compute, then the 75/25 split is specifically about physical-world AI, not about chat assistants. That is a strategic signal. Musk may have concluded that the next frontier of AI is not the language model but the autonomous system. He may also have concluded that OpenAI is winning the language-model war and he cannot catch up through general-purpose training alone. The only path to differentiation is to make AI physically manifest, in robots and spacecraft. TeraFab is the name for that bet. The 75 percent of compute allocation is not a technical preference for orbital inference. It is a declaration that the future of AI lies in taking models off the cloud and putting them into moving, armed, and dangerous physical objects. That declaration is exactly why the low relevance assigned to ethics is so uncomfortable. A spacecraft with onboard AI inference can be used for civilian earth observation, debris avoidance, and autonomous docking. The same architecture can be used for reconnaissance, orbital inspection, proximity operations, and kinetic collision. There is no technical marker in the GPU cluster that separates civilian autonomy from military capabilities. The report tries to dismiss this with a single line, but the security implications are the most consequential edge case in the entire project. If TeraFab is real, the United States federal government will not allow a private citizen to build 85 megawatts of orbital AI compute without significant oversight. The Committee on Foreign Investment in the United States, the Federal Communications Commission, the Federal Aviation Administration, and the Department of Defense will all have opinions. The report does not map the regulatory path. That omission is inexcusable for an infrastructure analysis. Let us now examine the energy economics from the perspective of a protocol audit. The report correctly states that 1TW in Texas would require nuclear power plants or very large gas plants. A single large nuclear reactor typically produces 1.6 to 2.0 gigawatts of electrical output. To supply 1,000 gigawatts, one would need approximately 500 large nuclear reactors, all exclusively devoted to TeraFab. The entire United States currently operates around 93 commercial reactors. The world operates about 440. The construction time for a single nuclear reactor in the United States has historically been a decade or more. There is no version of this mathematics that supports a plausible near-term TeraFab. Even the 114MW average-power reading would require a substantial dedicated energy supply, but it would be within the realm of a large industrial site with gas turbines or a grid connection. The report’s failure to distinguish between these two readings is not a minor editorial oversight. It is the central analytical failure of the source material. I have audited too many protocols where the founding team released a white paper with elegant mathematical notation and omitted the edge cases. The proof was always in the unverified edge cases. For TeraFab, the edge cases include the power contract, the land permit, the water supply, the optical backbone, and the control-plane architecture. None of these appear in the initial material. Instead, the material offers a narrative of inevitability. It is as if the project expects to be judged by its ambition rather than by its engineering artefacts. This brings me to the question of decentralization. My background is in Layer 2 research, and I have spent years warning that optimistic rollups are not an autonomous security layer but a delay in truth extraction. The sequencer is a single node. The fraud proof window is a deadline, not a freedom. Decentralized sequencing has been a PowerPoint slide for two years. TeraFab is the opposite of decentralisation. It is a vertically integrated, centrally governed, physically concentrated compute empire. That is not a critique in itself; many complex systems require centralisation at the control layer. But it is a warning. The source report treats vertical integration as strength. The history of infrastructure is full of vertically integrated empires that collapsed because the control layer could not process its own hidden faults. Consider the Ronin bridge again. Ronin did not fail because the chain was slow. Ronin did not fail because the smart contract code had a reentrancy bug. Ronin did not fail because of a random exploit. Ronin was engineered to trust a limited set of validator signatures, and the compromise of those signatures was enough to drain the bridge. The proof is in the unverified edge cases. The edge case was the number of validators. The edge case was the assumption that each validator key was held by a distinct and independent entity. The TeraFab narrative has the same shape. It asks us to trust that the power will be generated, that the spacecraft will be launched, that the robots will be built, and that the interlocking Musk companies will transfer value fairly. Each trust assumption is plausible in isolation. Together, they form a complex system with too many unverified edges. Complexity is not a shield; it is a trap. The more moving parts a system has, the more ways it can fail. The official TeraFab narrative hides the failure modes behind the sheer size of the ambition. Who audits the power purchase agreement? Who audits the satellite firmware update pipeline? Who audits the transfer price between Tesla and SpaceX? Nobody. There is no independent verification layer. There is no public cryptographic proof of compute allocation. There is no smart contract enforcing the 75/25 split. There is only a press narrative. In the crypto world, we would call this a centralized oracle with no slashing conditions. The report predicts that TeraFab’s existence will trigger a second round of the global compute race. The first round was about GPU supply. The second round will be about power capacity and physical-world deployment. I agree with that prediction, but I would add a caveat. The race will not be measured in terawatts. It will be measured in negotiated power contracts, committed capital expenditures, and operating licenses. The winner will not be the company that announces the largest number. The winner will be the company that can convert energy into reliable, secure, physically deployed AI systems without blowing up its balance sheet or its legal standing. TeraFab, as currently defined, is not even a contender in that race because it has not given us a date, a unit, or a legal entity. The source report’s investment analysis also needs scrutiny. It says the compute narrative will have a significant impact on the valuation of Musk-linked companies and related industrial themes. This is true in the short term. Financial markets are pattern-matching machines. If the phrase 1TW is attached to Musk, traders will immediately look for nuclear energy stocks, power infrastructure companies, and satellite manufacturers. But the valuation effect is not the same as the operational effect. A year from now, we will be able to measure TeraFab by a different set of signs: does the ERCOT interconnection queue contain a new application from a Musk affiliate? Has the Nuclear Regulatory Commission docket received a hearing request? Has a public filing with the SEC disclosed a capital expenditure commitment? The source report does not know. Without those signs, the investment thesis is a meme in search of a balance sheet. There is a deeper mathematical problem with the 75 percent allocation to spacecraft. Let us assume, for a moment, that TeraFab abandons the terawatt fantasy and settles for a real 114MW facility. The arithmetic of orbital compute is brutal. A typical modern GPU accelerator weighs tens of kilograms and requires hundreds of watts. To put 100 kilowatts of compute into orbit, you need to launch several tonnes of electronics, plus solar panels and thermal control. The cost of launch remains a barrier, although SpaceX has reduced it dramatically. Even if you use Starlink as a ride-share platform, the mass-to-orbit constraint is real. A constellation of 10,000 satellites with 100 watts of inference compute each gives you exactly 1 megawatt of total orbital inference. To get to 85 megawatts of orbital AI compute, you need 850,000 satellites, or a new class of massive orbital data centers. The source report’s predicted transition from communication network to space edge computing network is directionally correct, but the scale is off by roughly three orders of magnitude. That is not a detail. That is the difference between a satellite startup and a spacefaring civilisation. The machine intelligence that operates orbital infrastructure will likely be much smaller in footprint per capability than terrestrial GPUs. Neuromorphic chips, radiation-tolerant FPGAs, and application-specific inference accelerators will be the workhorses of space AI. These chips are not interchangeable with H100s. The software stack is completely different. The training paradigm is constrained by the need to freeze models and validate them rigorously before launch. The loop for updating a spacecraft model is measured in days, not milliseconds. A terrestrial data center can be reconfigured at will. An orbital fleet cannot be touched. The report does not engage with any of these architectural realities. It treats AI spacecraft as if they were simply servers in the sky. That is the same error that every Layer 2 marketing deck makes. It promises that the user experience of a rollup will be identical to a base layer, with higher throughput and lower fees, while ignoring the fact that the fraud proof window, the sequencer liveness, and the data availability guarantee create a very different trust profile. The user does not notice the difference until the sequencer stops. The spacecraft operator will not notice the difference until the satellite goes silent. TeraFab is a Layer 2 in disguise. It is a centralised compute layer that promises to deliver intelligence to the physical world, but it delays the truth of its own architecture. I need to be clear about what I am not saying. I am not saying that Musk will never build a large-scale compute facility. I am not saying that autonomous robots and spacecraft are impossible. I am saying that the report’s central claim, 1TW, is not a meaningful engineering specification. It is a rhetorical device. As a forensic analyst, I am trained to separate the rhetorical layer from the protocol layer. The protocol layer of TeraFab contains no code. It contains no public specification. It contains no reproducible benchmark. It contains nothing that can be independently verified. When I wrote my stress-test report on Solana’s TPU path, I included a reproducibility section so other researchers could run the same test against the same validator network. I published the code. I showed the raw latency curves. I did not ask readers to take my word for it. The source report on TeraFab does none of that. It asks readers to accept a 1TW number with no definition, a 75/25 split with no allocation mechanism, and a policy conclusion with no regulatory analysis. For a report that claims to be a deep dive, it is remarkably shallow at exactly the point that matters: the accuracy of the primary claim. There is one more missing piece. The source report mentions that TeraFab may be a shared compute platform for Tesla, SpaceX, xAI, and X. If that is true, then the 25 percent allocation to Optimus is not a binding commitment. It is a negotiation. The allocation could change based on which company has the strongest board representation. Tesla is a public company with fiduciary duties to outside shareholders. SpaceX is privately held. xAI is a separate venture. X is facing debt obligations and a chaotic advertising environment. Forcing these entities to share one compute fabric requires a governance layer that has not been specified. Who decides the allocation? Who audits the usage? Who benefits from the transfer price? The report calls this an internal arsenal. I call it an unverified oracle. Let me make a prediction. Within the next twelve months, there will be announcements about Musk-affiliated compute capacity. They will not mention TeraFab as a single 1TW project. They will mention a new data center campus in Texas, a nuclear power purchase agreement, a cluster of GPUs at a specific tower in Memphis or Austin, and a software stack that connects Starlink satellites to a ground-based inference engine. The TeraFab name may fade. The 1TW target will be quietly abandoned or reinterpreted as an aspirational roadmap over multiple decades. The physical-world AI thesis will remain. The 75 percent to spacecraft may be adjusted downward when the FCC and the Pentagon ask pointed questions. The deeper takeaway is not about TeraFab. It is about the way we evaluate infrastructure claims. In the blockchain world, we learned that a white paper is not proof. A GitHub repository is not proof. A testnet is not proof. The proof is in the sustained behaviour of the system under adversarial pressure. The proof is in the unverified edge cases. TeraFab is a test of whether we have learned that lesson outside the blockchain bubble. The financial press will cover TeraFab as a story about scale. The correct coverage will treat it as a story about missing data. Complexity is not a shield; it is a trap. TeraFab’s complexity is carefully designed to make the audience feel that a 1TW installation is only a matter of will. It is not. A terawatt requires energy policy, electrical transmission, heat rejection, and physical security. It requires a workforce of tens of thousands. It requires a supply chain for transformers, switchgear, and cooling towers that does not currently exist at the required volume. The report treats these as trivial logistics. They are the core problem. When the math holds but the incentives break, the system fails. With a single-entity project, the math can be simplified: one owner, one budget, one operating plan. With TeraFab, the ownership is shared across at least four entities with different shareholder bases, different risk appetites, and different regulatory exposures. The math of compute allocation may hold on paper. The incentives will not hold under fiscal pressure. Tesla’s board cannot easily justify sending a quarter of a giant compute facility to robotics if SpaceX is not paying market rates for the other three quarters. The incentives are fragile. That is the architectural weakness. Layer 2 is merely a delay in truth extraction. TeraFab is a delay in power extraction. It is a promise that the future will be built before the present is audited. The truth, when it emerges, will be found in the ERCOT queue, in the nuclear licensing docket, in the SEC filings, and in the discrepancy between what was announced and what was built. I started this analysis with the line that silence in the slasher was the first warning sign. TeraFab’s silence is not about a missing horror-movie soundtrack. It is the silence of a project that has no date, no unit, no legal entity, and no accountability. That silence is the warning. I do not need to decide whether TeraFab is a fraud. A project can be completely sincere and still be completely unverified. I do not need to decide whether Musk can build autonomous robots and spacecraft. He has already proven that he can build hard physical systems at scale. The question is whether TeraFab, as reported, meets the evidential standard of an engineering roadmap. It does not. The next phase of the AI compute race will be decided by energy and trust. Energy is the visible constraint. Trust is the invisible one. TeraFab asks us to trust that a terawatt can be conjured, that interlocking companies will cooperate, and that autonomous spacecraft will be controlled responsibly. I am not willing to extend that trust on the basis of a number with no timestamp. The proof is in the unverified edge cases, and TeraFab is an edge case the size of a continent.