VaticinusForecasting where
scarcity moves next
We track filings, permits, trade flows, and research output to find the physical input that stops scaling before the market prices it.
Every call is dated, carries a kill-criterion, and is scored in public. Across 793 resolved questions, the system beat the fair early crowd at 95% significance.
Early reads, nothing scored yet
Where the constraint is already moving our way
We called what the energy and chip booms are now hitting: the binding constraint sits one layer upstream, on inputs that cannot scale on demand. Nothing here is scored yet, but on dated price anchors each constraint has already moved the way we said.
We traced the AI buildout's constraint past the GPU into two physically capped metals: hafnium, hostage to three inelastic buyers at once, and ruthenium, the interconnect metal that takes over below 2nm.
The contrarian call: Chinese polysilicon stays below cash cost through 2027, because regulators banned the coordinated supply cut the market had already priced in.
The same ledger keeps what we got wrong. It logs AI compute as a real miss we would have caught late, and flags GLP-1 and the EUV step as gaps it cannot adjudicate, never quietly scored as wins.
Open calls
Seven bottleneck calls, scored when they come due
Seven live bottleneck calls from the sealed ledger, June and July 2026, each fixed at publish: probability, resolution date, kill-criterion. The July batch also survived an adversarial pass, with independent skeptics trying to kill it on live data. An honest 30% counts for more here than a defended 50%.
AI firm power is rationed by OEM manufacturing slots, not generation economics. Our call: the quoted delivery lead-time for a new heavy-duty F/H-class gas turbine (100 MW and up) is still 3 years or more at end-2027, with the next open OEM slot sitting in 2030 or later.
The binding constraint on new US power is the process, not the hardware: the interconnection queue and transmission. Our call: the median queue duration for completed projects, on Berkeley Lab’s own series, is still 4 years or more at end-2028.
Humanoid-robot demand is the loudest new story in heavy rare earths. Our call is that it stays a story for now: no humanoid OEM signs a binding, publicly disclosed ex-China Dy/Tb offtake at a premium to Chinese metal before 2027-06-30.
The electrification bottleneck has migrated off the equipment and onto the people who install it. Our call, conditional on trades labour staying the pace-setter: the electrician wage premium over equipment PPI persists or widens through 2028.
ITU "first-come" spectrum priority looks like a settled land grab, but deployment deadlines quietly make it use-it-or-lose-it. Our call: by end-2030 a majority of pre-2024 non-GSO Ku/Ka filings miss the first milestone, and protected priority concentrates in five or fewer systems.
The bottleneck under "data-center power" is moving onto a narrower asset: a firm, transferable right to actually energize a large load, which FERC's December 2025 co-location order starts minting. Our call: by end-2027 PJM runs a live firm-demand interconnection product and sites holding an executed large-load agreement trade above twice comparable unpowered land.
The DRC's export cap throttles about three quarters of mined cobalt; an ex-DRC refined stockpile is cushioning the deficit for now. Our call is the second leg: once that buffer empties, the monthly average clears $75,000/t, with no traded forward pricing it.
Research notes
The notes the calls came from
Each note names one constraint, the mechanism behind it, and the calls it produced, fixed at publish. Some resolve within a year; the long-horizon ones run out to 2033. Open the PDF and check the reasoning yourself.
Announced megawatts are easy to claim. The scarce asset in AI infrastructure is source-grade proof that a named campus can actually energize on the claimed schedule.
At least three public large-load projects get pushed, conditioned, contested, or demoted in official sources.
Why SanDisk, Micron, and the memory complex went vertical: AI demand has moved from GPUs into NAND, HBM, enterprise SSDs, controller firmware, and prepayment-backed allocation.
Enterprise SSD allocation becomes the first storage gate AI operators feel.
The lead note: upstream inputs fixed by physics, treaty, or demography while the system above them keeps scaling. The rent lands on what cannot answer a price signal.
Hafnium has no ore. Three industries now collide on one byproduct trickle.
The constraint often sits below ASML and the fab shells: the artisanal sub-step, the process engineer, the certified welder, and the isotope cascade.
Sub-5nm ramp veterans are the leading-edge input capital cannot clone.
what vati does
Forecasting by judgement, graded in public
The best forecasters in messy domains are still humans who reason from evidence, not curve fits. Vati is built to forecast the same way, at machine speed, and every call gets graded in public.
Our preprint shows why the field has been measuring forecasting skill wrong. Across 25 leak-free ForecastBench rounds (2024 to 2026), frontier LLMs add almost no information beyond the market price, and more scale has not changed that. Human superforecasters clear the same bar by an order of magnitude.
What is Vaticinus?
Vaticinus is an independent forecasting lab. We figure out where scarcity and value are headed, write the call down before anyone knows the answer, and let it get scored in public.
Every call rests on one bet: rent goes to whatever input can't scale when demand arrives. So we look for the bottleneck a layer under the headline, check whether the market has already priced it, and put a date and a number on what would prove us wrong.
Vati has an edge on exactly the topics where a purely data-driven approach falls short, the ones where judgement decides the answer and data alone cannot.
We've developed our system to make medium term predictions (1 week - 1 year out) about geopolitics, business, policy, technology, and culture.
Performance
We compete where someone else keeps score
Most AI forecasting lives in screenshots. Ours lives on the field's own scoreboards, under one name anyone can check. We run a live public record on Metaculus across 13 tournaments, we are in ForecastBench's blind, leak-free round and built to land #1 among bots, and we published the analysis that shows the field has been measuring forecasting skill wrong. None of it is a finished track record, by design: every number is scored by someone other than us, when the questions close. That is the whole point.
Where we already compete, in public:
Leak-free calibration, proven now
- 793 questions that already resolved, forecast by models that could not have known the outcome
- Beats the fair early crowd at 95% significance across random, time, and domain splits
- Sealed and pre-registered, so nothing is fit after the fact. No waiting to check it.
- AI & tech
- Politics & geopolitics
- Macro & markets
- Crypto, science, business
ForecastBench - dataset half
- In the blind, leak-free 21 June 2026 round, built to land #1 among bots
- Scored by ForecastBench, not by us; the number posts when the round resolves
- The same dataset half where the best bots already clear the superforecaster bar
- Geopolitics
- Elections
- Technology
- Energy
Beyond Brier - we set the measure
- Our preprint shows frontier LLMs add almost no information beyond the market price
- Across 25 leak-free rounds, and more scale has not closed the gap
- Human superforecasters clear the same bar by an order of magnitude, which is why we forecast by judgement
- ForecastBench, 2024 to 2026
- 25 leak-free rounds
- LLMs vs superforecasters
The score is downstream. The work starts with the evidence: what the system checks, what should move first, and what would make a thesis wrong.
The data layer
The evidence we check before a call
We track the parts of the world that move before consensus does: research that flags a capability, trade flows that reveal who controls a chokepoint, permits that show what is getting built, and policy that moves the constraint.
Research
Trade flows
Policy
Buildout
Independent
A small Berlin research lab with no fund, platform, or house book steering the calls.
Leak-free
Calls are dated and sealed before the clock starts, then graded only on what happens after.
Scored outside
We compete on ForecastBench, Metaculus, and the public record, where someone else keeps score.
Work with us
Dated calls on the constraints in your book.
Built for desks that own these questions: sector funds in the fuel cycle, critical minerals, and grid power; pod PMs and sector analysts; commodity research desks; family offices with resource books. If uranium conversion, rare-earth magnets, grid steel, or power semiconductors sit in your book, there is already a dated call on the record about it. For the positions you size, we point the engine at your coverage: where the binding constraint is moving, whether the move is priced, and a probability with an interval, a resolution date, and the one thing that would prove it wrong.
Founding subscribers get the running record of dated calls, a quarterly deep dive on their sector, a chat seat, and a direct line: $5,000 a year, locked at the founding rate. The record is young and we say so. Every call is sealed at publish and scored in public when it comes due, so what you judge is the record, not the pitch, and the price rises as the record resolves. The chat is free and always on if you want to test the engine first.
Operators get a different door. If a campus, sourcing, or qualification decision hangs on power, permits, equipment, or export permission, we run a 10-15 day decision sprint: map the serial gates between you and the deadline, grade which ones are evidenced and which are assumed, and set the dated signals and kill conditions before capital commits. Sanitized specimens available on request.
Grab a 30-minute call
Pick a time below, or email [email protected].
