AI can already predict hurricanes, cyberattacks and world events. A new generation of companies is trying to turn the future itself into data.
For most of human history, predicting the future belonged to prophets, gamblers, economists, meteorologists and the occasionally lucky uncle who swore he knew which stock was about to explode.
Artificial intelligence is beginning to change that.
Not because a machine has suddenly developed supernatural powers. It hasn’t.
Something potentially more consequential is happening instead: computers are becoming extraordinarily good at gathering enormous amounts of information about the world, finding patterns humans cannot see, assigning probabilities to future events and continuously correcting themselves as reality unfolds.
The result isn’t a crystal ball.
It is a probability machine.
And an entire industry is beginning to form around it.
For Charlotte, this should be more than an interesting technology story. This is a city whose economy revolves around banking, energy, transportation, logistics, construction and corporate decision-making — industries where knowing what happens next even slightly earlier can be worth enormous amounts of money.
The race is now on to build machines capable of doing exactly that.
The AI superforecasters
The most direct attempt may be coming from a London-based startup called Mantic.
Its mission is unusually straightforward: build machines capable of predicting messy real-world events better than expert humans.
That used to sound like Silicon Valley marketing.
Then the machines started winning.
During the Summer 2026 Metaculus Cup — a forecasting competition involving 885 forecasters making more than 88,000 forecasts across 58 questions — AI systems finished first and second overall. ManticAI placed second, ahead of every human participant. FutureSearch, another AI forecasting company, finished fifth. (Metaculus)
The questions weren’t simply math problems with hidden answers.
They concerned events that genuinely had not happened yet.
Forecasters had to estimate things such as economic indicators, technology developments, public-health outcomes, business events and geopolitical developments. They were then scored after reality provided the answer.
That distinction matters.
You cannot memorize tomorrow’s news from a training set.
Reuters reported in September that Mantic raised a $25 million seed round led by Radical Ventures, with participation from Microsoft’s M12, Thinking Machines Lab and Balderton Capital. Companies and government agencies are already experimenting with the technology. (Reuters)
Mantic is targeting roughly the one-week-to-one-year range across business, technology, policy, geopolitics and culture. (Mantic)
In other words, it isn’t merely asking AI what happened.
It is asking AI what happens next.
Building an actual model of the world
Then there is FutureSearch.
The company was founded in 2023 by former Metaculus and Google employees explicitly to build AI capable of predicting the future. (FutureSearch)
Its most interesting experiment goes beyond answering individual forecasting questions.
FutureSearch has begun building what it calls a shared world model.
Suppose an AI is separately asked about oil prices, electric-vehicle adoption, an international conflict, inflation and airline profitability.
Those questions aren’t independent.
War can affect oil. Oil affects transportation. Transportation costs affect inflation. Interest rates affect consumers. Consumer behavior affects car purchases and travel.
Traditional forecasting models often treat these problems separately.
FutureSearch is attempting to create a network of interlocking probabilities so that information learned while answering one question can affect forecasts elsewhere.
The company says its system now draws from thousands of forecasts, identifies shared drivers among them and reconciles new predictions with its existing picture of the world. Internal testing found a measurable improvement in forecasting performance after adding the system, although live independent benchmarks remain the more important test. (FutureSearch)
As of September, FutureSearch reported a No. 2 position among nearly 300 systems in the live Metaculus FutureEval tournament and has begun publishing real-money prediction-market results as another test of its models. (FutureSearch)
This is getting remarkably close to the science-fiction version of forecasting.
Not knowing the future.
Modeling the world’s possible futures and constantly changing their probabilities as new evidence arrives.
The company watching a million signals at once
If Mantic and FutureSearch are trying to forecast tomorrow, Dataminr is trying to identify what is beginning to happen right now before most of the world realizes it.
Its systems analyze more than 1 million public data sources spanning text, images, video, audio and sensor information in more than 150 languages.
Dataminr says its infrastructure processes more than 43 terabytes of information each day and identifies nearly half a million events, risks and threats. (Dataminr)
Imagine thousands of seemingly unrelated signals suddenly appearing:
a local social-media post,
a government notice,
a strange satellite observation,
a video,
shipping information,
emergency communications,
an obscure regional news story.
No human newsroom, bank, government or corporate intelligence department could continuously process all of it.
AI can.
Dataminr’s newer systems are moving from merely detecting events toward what the company calls near-term predictive intelligence — attempting to anticipate how an event may develop after the first signals appear. (Dataminr)
That is a subtle but enormously important transition.
The valuable question is no longer simply:
What happened?
It becomes:
Something appears to be starting. What is it likely to become?
The future from space
Some of the best predictive AI is looking down at Earth from orbit.
BlackSky operates satellites capable of repeatedly observing locations and feeding imagery into AI systems that identify changes on the ground. Its technology is used for rapidly unfolding geopolitical, infrastructure and security events. (BlackSky)
Spire Global operates a satellite constellation collecting information about weather, aircraft and ships. The company’s systems combine those observations with machine learning to forecast weather conditions, voyage risks and other operational outcomes. In 2026, Spire said one of its new generative-AI weather systems outperformed ECMWF’s sub-seasonal forecasts in its testing, particularly several weeks into the future. (Spire Global)
This points toward an important feature of future AI systems:
They won’t rely only on the internet.
They will have sensors.
Satellites, cameras, radar, aircraft, ships, financial transactions and other machines increasingly create a continuously updating digital representation of the physical world.
AI gets to analyze it.
Weather may be the proof that this works
The strongest scientific demonstration of AI forecasting may already be weather.
Google DeepMind’s GenCast produces probabilistic forecasts up to 15 days into the future.
In testing published with its research, GenCast outperformed the European Centre for Medium-Range Weather Forecasts’ leading ensemble system on 97.2% of the evaluated targets — and 99.8% of targets more than 36 hours into the forecast. (Google DeepMind)
Meanwhile, Boston-based Tomorrow.io has built its own commercial weather satellite constellation and combines proprietary observations with AI forecasting models.
Its goal isn’t merely telling you whether it rains.
The company wants an airline, trucking company, military unit, insurer or energy provider to know how weather will affect its operations before the consequences arrive. Tomorrow.io says its first satellite constellation has achieved roughly hourly global revisit capability, while its next generation is being designed around AI-native sensing and forecasting. (Tomorrow.io)
Weather offers an important lesson for the rest of AI forecasting.
The machine does not need to predict the future perfectly to be immensely valuable.
It simply needs to predict it better than the alternatives.
Cybersecurity has its own crystal ball
The same concept is playing out in cybersecurity.
Recorded Future continuously collects information from across the open internet, technical infrastructure, dark-web sources and other datasets to identify emerging threats before they reach a company’s network.
Its basic premise has always been predictive: indicators of tomorrow’s cyberattack often exist somewhere in today’s data.
Mastercard considered that capability valuable enough to acquire Recorded Future for $2.65 billion in 2024. The company serves more than 1,900 customers globally. (Reuters)
Again, the competitive advantage is time.
Discovering an attack after it occurs is useful.
Discovering the conditions that suggest an attack may be coming is far more valuable.
And then there is Palantir
Palantir occupies a slightly different position.
It isn’t primarily trying to become an all-knowing forecaster of world events.
It is building something potentially just as important: the system that connects predictions to decisions.
Palantir’s software can pull together an organization’s supply-chain data, inventory, factories, transportation systems, customers and external information. Forecasting models can then be attached to what Palantir calls its Ontology — effectively a living digital representation of how an organization operates.
Its software supports simulations that ask what happens under different assumptions before an organization makes the real-world decision. (Palantir)
That closes the loop.
Observe the world. Predict what may happen. Simulate the alternatives. Choose an action. Observe what actually happened. Learn. Repeat.
That feedback cycle may eventually matter more than the AI model itself.
Metaculus: the scoreboard
There is one company worth watching that isn’t trying to sell the crystal ball.
It is trying to determine whether the crystal ball actually works.
Metaculus operates forecasting tournaments and FutureEval, a benchmark specifically designed to measure whether AI systems can accurately predict unresolved real-world events.
Its database now contains millions of predictions across thousands of questions. (Metaculus)
And it reveals an important reality behind all the hype.
Simply asking a powerful chatbot what will happen is still not the same thing as building an excellent forecasting system.
Metaculus’ current benchmark shows professional human forecasters still outperforming standalone frontier AI models. But specialized AI systems that combine frontier models with research, calibration, multiple agents and forecasting architecture have advanced much further. (Metaculus)
That is why the Summer 2026 result was so significant.
A machine finally beat the humans in a live forecasting competition.
Not because it knew the future.
Because it was better at estimating uncertainty.
The real opportunity
The company that eventually becomes the world’s most powerful forecasting engine may not exist yet.
The ingredients, however, are appearing everywhere.
Dataminr sees public signals.
BlackSky and Spire watch the planet.
Tomorrow.io and DeepMind model physical systems.
Recorded Future watches digital threats.
Mantic and FutureSearch reason about uncertain events.
Palantir connects models to actual organizations.
Metaculus measures whether any of them are actually good at predicting anything.
Now imagine those capabilities converging.
A machine constantly observing economic data, scientific research, satellite imagery, supply chains, weather, financial markets, corporate filings, public behavior and millions of other signals.
Not to produce one grand declaration about what will happen.
To maintain millions of probabilities:
35 percent. 62 percent. 8 percent. 81 percent.
And then update them every time reality changes.
For Charlotte’s bankers, entrepreneurs, investors, energy companies, logistics businesses and corporate headquarters, that could eventually create a new kind of competitive advantage.
The most valuable AI may not be the machine that writes the fastest email.
It may be the machine that notices something changing in the world before everyone else does.
For thousands of years, humans have wanted to see the future.
We may finally be building something stranger:
machines that can’t see tomorrow — but can calculate it better than we can.
