Byrne Hobart
In this issue:
- The Finance of Alchemy—A fusion startup claims to have found a way to turn mercury into gold. Aside from being another victory for the "always bet on tech ideas from Neal Stephenson novels" thesis, how big a deal would this be? If you can produce much more of something that derives its value from scarcity, what do you do about it?
- SPACs, Back—SPACs were too profitable to stay disreputable in 2021, and even though that cycle burnished their reputation for dubious deals, they're still working.
- Sanctions and Supply Chains—China and Nvidia walk back expectations on H20s.
- AI Filtering—If you can't easily afford to bet big on AI, you have the opportunity to bet against it instead.
- Measurement Errors—The impact of tariffs shows up in the intangibles.
- Data Labeling—Turning white-collar work from COGS and opex to capex.
The Finance of Alchemy
There's a surprising range of companies where their unit economics for any given start-to-finish sale process are negative, but they can make money in the aggregate by sharing costs. Whether you're a mid-century slaughterhouse that breaks even on pork but makes a small positive margin after selling insulin from pig pancreases, or a quant hedge fund that loses money after costs on every single signal, but that can combine multiple signals into something that turns a profit, or a streaming service that can't make money with pure subscriptions or pure ads but can make money doing both—it's a common pattern, and one that leads to counterintuitive businesses with durable competitive advantages.[1]
Here's a fun one: fusion has always famously been the big-thing-after-the-next-big-thing, a technology that's perpetually ten years away from viability, etc. The entire energy industry is an energy-in/energy-out proposition, which started with converting human and animal labor into wood and peat, and has progressed to converting energy from hydrocarbons into either renewables or more hydrocarbons. Fusion is like that, just with a negative ROI: you can turn a lot of non-fusion energy into a bit less. (Annoyingly, the fusion gain factor is quoted in terms of total energy output from the reaction divided by total energy input, but given that none of the processes involved (particularly in how the heat from the output is turned into useful energy) are perfectly efficient, the breakeven ratio varies depending on how broadly you scope the question, with the narrowest being scientific breakeven (the fusion reaction produces heat as quickly as the device dissipates it) to economic or commercial breakeven, i.e. a fusion power source whose revenue exceeds its costs.[2]
Marathon Fusion, a fusion startup, recently made a surprising announcement (technical details here) about an unusual potential revenue source: alchemy. The backstory: your reactor needs a blanket of some element that will slow down neutrons. As these neutrons encounter atoms of whatever that element happens to be, they turn it into a potentially-unstable isotope of that element, and that isotope decays into something else. It turns out that one of the viable isotopes to choose for this is 197Hg, which decays into 197Au. Which is a lot more interesting if you paraphrase it as "fusion reactors can turn enriched mercury into slightly radioactive gold, which, after 6.8 years, will be in the least hazardous nuclear waste category. (If you wait 17.7 years, a troy ounce of gold produced this way is less radioactive than a banana and doesn’t require any labeling.)[3] Of course, there's a lot of ground to cover between "this works in theory" and actual gold production. But it's interesting to consider the hypothetical.
Since there's more than enough gold to meet industrial demand, at least for now, the relevant question is about financial demand. For financial purposes, radioactive gold is worth about as much as the regular kind, since it can perform the core function of sitting in a vault somewhere, backing some tradable electronic representation of gold (a futures contract, a unit of the GLD ETF, etc.).[4] (If anything, it's arguably more valuable than regular gold, because it's more inconvenient to steal.)
The paper gives a wide range for the important question of how much gold could actually be produced this way. 1.5 gigawatts of thermal output would produce 450-600 MW at the same conversion efficiency as fission power. This level of output would require 100-400 tons of mercury in a magnetic confinement fusion device, which could produce three tons of gold per year. At current wholesale prices, 525MW of power produces about $266m of revenue, assuming high but not 100% utilization. The gold component is $320m. So, all else being equal, fusion can produce around 120% more revenue per megawatt of capacity through its sideline of alchemy.
All else is clearly not equal. Gold is interesting for two reasons: it's a helpful industrial metal, very ductile, dense enough for infrared shielding (as on astronaut's helmets), resistant to corrosion and oxidation, etc. It's often the best choice when money is no object, and the fact that it still has industrial uses even with a high price driven by speculative demand indicates that there are some uses for which there really is no substitute. The second reason that gold is interesting is that it's one of the longest-established Schelling Points for a universal intermediate to exchange different goods and services. It's been viewed as money for longer than any other medium (it’s outlasted livestock and tends to be the top store-of-value choice when silver is an otherwise good option), and if you dig up some gold object buried two thousand years ago, no matter how beaten-up and deformed it's gotten, it's retained its purchasing power better than any other investment you could have made (though, of course, not for the person who buried it—coin collectors are also collecting talismans of bad luck, probably during the sort of incident that leads to other assets losing all of their value, too). The reason it worked so well as money is that the supply is inelastic; if your currency is cowrie shells, chunks of iron, beads, packs of cigarettes, etc., demand for savings will eventually produce a higher supply of the product. Paper money runs that risk, too, though the right institutions run by the right people can keep that devaluation in check. Gold is portable enough for flee-the-country-with-one-suitcase type purposes, though it gets tricky with nation-state levels of gold; Iraq seized Kuwait's gold in the Gulf War, the Nazis captured about $36bn worth at present prices from countries they conquered, and the USSR got slightly more gold from Spain during the Spanish Civil War.[5] Its supply is inelastic, and annual production runs at a little under 2% of all the gold in existence. Once there's another source of gold, does that mean gold doesn't work as a Schelling Point for inflation-resistant saving? If so, where does that money flow? Will crypto true believers start massively subsidizing the fusion industry in order to eliminate the one big competitor in the "quasi-currency with low-and-predictable inflation" niche? Or will gold producers decide to beat them to the punch by pouring money into quantum computing?
The easy answer here is that the swing producer has an incentive to moderate supply in order to reduce price swings, because they'll make more money over time, and they're in it for the long haul. At various points in the history of OPEC, Saudi Arabia has exercised either iron discipline or a crazy risk tolerance in its production policies, occasionally trying to punish irresponsible overproduction by flooding the market so everyone else loses money for a while until supply can equilibrate. Someone who hypothetically can make arbitrary amounts of gold will be able to convert that into many more dollars over time if gold traders know exactly how much incremental gold there will be (and, to be fair, it takes a lot of fusion . Most currencies are designed to slowly depreciate over time, which tends to keep economies more fluid that stable price levels would. Gold is one of a handful that doesn't. But even if it does, it's an incremental change in the dilution that's already happening. Being bullish on gold is a bet that demand will rise by an amount that offsets the increase in supply from mining, and given that gold is big enough to count as an asset class of its own, there's obviously a large investor population willing to tolerate some dilution.
This makes fusion superior to asteroid mining in the ranks of "thus-far hypothetical ways to acquire an immense quantity of gold" business. If fusion can really perform economically useful alchemy at scale, it won't be the end of the gold business unless significant mistakes are made, but it will definitely redistribute wealth within it. And while it's annoying to make an inflation-resistant asset a bit less inflation-resistant, it ends up serving the same purpose that inflation does in a fiat model—redistributing wealth towards public goods, such as economically-viable fusion power.
Thanks to the Marathon team for answering my question. All remaining technical errors are a) mine, and b) exist despite their best efforts.
This is easiest to see in the quant example: if firms use multiple signals that make money before transaction costs but lose money after those costs, then an independent operator who finds just one good signal doesn't necessarily know if they'll find enough of them to outpace transaction costs. Whereas an already-profitable company can underwrite that risk. It's still possible for the industry to fragment, but one way this can happen is that the bigger companies limit how much they use some signals, to ensure that they aren't too big in a crowded trade. But it's hard for a secretive industry to really know in the aggregate how crowded a trade is, though they can put a lot of energy into guessing. If that means there's capacity that can be absorbed by new entrants, it also means that those new entrants can blow up all at once. This roughly describes what happened in 1998 and 2007-8. New entrants need to scale fast enough to survive a crisis without really knowing whether they'll be surviving or merely contributing to the next crisis. ↩︎
You'll know we're closer to viable fusion when people talk about unit-profitable breakeven versus ROIC >= WACC breakeven. ↩︎
If there does turn out to be some difference in the value of temporarily-radioactive and non-radioactive gold, the good news is that we have lots of data on this question thanks to crypto, where market participants sometimes have to consider the relative value of locked or unlocked tokens. ↩︎
One way to think about this is that gold, as a financial asset, is a gold-denominated zero-coupon bond that can be redeemed at any time. And radioactive gold is the same thing, but with redemption available a few years in the future. If you happen to need to deliver gold, and the only kind you have is radioactive, that's a problem, but there's plenty of other gold out there which you could buy to make that delivery. ↩︎
This gold was, depending on how you read the situation, either entrusted to Russia for safekeeping by the Republicans, who ultimately lost but who probably wanted the money to fund international communism regardless, or was used by the Republicans to hire a mercenary army (an especially distressing prospect for orthodox Marxists, since it involves an industrialized capitalist country backsliding into feudal war norms rather than advancing to a glorious socialist future). ↩︎
You're on the free list for The Diff! Last week, paying subscribers read about how much socializing will disappear thanks to chatbots ($), how the UK made the world's most catastrophic derivatives bet in the form of the triple lock ($), and thoughts on quasi-acqui-hires and iterated games ($). Upgrade today for full access.
Diff Jobs
Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:
- A leading AI transformation & PE investment firm (think private equity meets Palantir) that’s been focused on investing in and transforming businesses with AI long before ChatGPT (100+ successful portfolio company AI transformations since 2019) is hiring Associates, VPs, and Principals to lead AI transformations at portfolio companies starting from investment underwriting through AI deployment. If you’re a generalist with a technical degree (e.g., CS/EE/Engineering/Math) or comparable experience and deal/client-facing experience in top-tier consulting, product management, PE, IB, etc. this is for you. (Remote)
- A blockchain company that’s building solutions at the limits of distributed systems and driving 10x performance improvements over other widely adopted L1s is looking for an entrepreneur in residence to spearhead (prototype, launch, grow) application layer projects on their hyper-performant L1 blockchain. Expertise in React/React Native required. Experience as a builder/founder with 5–10 years in consumer tech, gaming, fintech, or crypto preferred. (SF)
- A hyper-growth company that was using ML/AI to improve software development/systems engineering before it was cool is looking for a product marketing manager to articulate their value proposition and drive developer adoption. If you started your career in backend engineering or technical product management, but have since transitioned (or want to transition) into a product marketing seat, this is for you. (Washington DC area)
- A Series B startup building regulatory AI agents to help automate compliance for companies in highly regulated industries is looking for legal engineers with financial regulatory experience (SEC, FINRA marketing review, Reg Z, UDAAP). JD required; top law firm experience preferred. (NYC)
- Deerfield-backed Series A startup building agents for healthcare administration (prior authorization, eligibility checks, patient scheduling) is looking for a senior software/AI engineer to build backend services and LLM agents. Experience building and monitoring production-quality ML and AI systems preferred. (NYC, Hybrid)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.
If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.
Elsewhere
SPACs, Back
SPACs are having their biggest year since 2022. 2021 remains a massive pull-forward year, where every remotely IPO-able company needed a good excuse not to go public through a SPAC that was eager to have them. So it's hard to believe that the last four years were enough time to grow a new crop of ready-to-go-public companies. Instead, what's probably happening is that enough investment banks and sponsors got in on the SPAC boom that it's hard for them to treat it as disreputable now. SPACs still work for companies that are in a rush to go public, however adversely-selected that cohort might be. So they make sense as both a viable financial product and one subject to a lot of opportunism.
Sanctions and Supply Chains
One of the big tech stories last week was the US's removal of restrictions on Chinese imports of advanced GPUs. Then Xi Jinping promptly warned Chinese officials not to overinvest in AI ($, FT), and Nvidia says it's already shifted production to other chips and will only be selling H20s out of existing inventory ($, The Information). It's very strange to see both parties to a trade deal talk down how much demand there is and how much supply is available to meet it.
Disclosure: Long NVDA, TSM.
AI Filtering
One feature of the AI business is that its capital-intensity means that big companies can choose to bet on it without seriously threatening their long-term viability, but smaller ones basically can't participate in the entire value chain and thus risk getting disintermediated. Which means that for some of them, the optimal move is to bet slightly against AI: DuckDuckGo is offering the option to hide AI-generated results. This is a feature that their larger competitors could implement, but their incentive is for their customers to not think much about AI specifically, and instead just to notice that the product's slightly better. It's a form of better that's hard to measure and prone to overfitting, but generally more efficient at transmitting information fast. But users vary in what they're maximizing for, even if part of what they want is, implicitly, to only look at things that someone thought were worth the time and effort to make.
Measurement Errors
One of the difficulties with measuring inflation is that you have to determine what someone's consumption basket is, but that basket can change based on the options presented to them. A good example of this is this WSJ piece that looks at how Amazon is raising prices on many low-price goods ($), which Amazon argues is misleading because they're cutting prices on other products, and which the Journal also notes is a slightly unfair comparison because other retailers, who didn't raise prices as much, were more likely to note that products were out of stock. If Amazon's costs are higher, they will of course need to adjust, and for now they might be raising prices to slightly ration their pre-tariff inventory rather than as a permanent change. These price increases can also slightly shift consumer spending towards higher-margin items where tariffs are a comparatively smaller contributor to the cost. Putting all of this together, some of the impact of tariffs shows up in actual price levels, but some unknown amount of it takes the form of reduced convenience and reduced selection.
Disclosure: Long AMZN.
Data Labeling
In the early days of AI, data labeling was mostly a simplistic task that could be outsourced to any cheap labor market with a large population that could speak a language representing a big chunk of global GDP, like English. Over time, tasks like "identify the stop signs in this image" or "read this text" have gotten more automated, and the tasks outsourced to data labelers have gotten more complicated ($, FT). There will be a period in the deployment of AI where the nature of white-collar tasks changes a lot less than the economics that drive demand for them: instead of performing semi-repetitive but skill-driven work to create revenue, they'll be doing the same work to create a data asset that later produces higher-margin, more scalable revenue. That shift from being opex/COGS to capex can temporarily make some workers' time more valuable, but if you're working for a data labeling business, you always need to be aware that the tasks you're doing for them are tasks that they expect to be automated in the near future. So if they're the high bidder for your skills, those skills are about to drop in value.