Summary

DeepOptica is not betting on “where to drill next,” but on answering in advance how much copper and gold lie underground—and what an entire mine is worth. 相子恒 wants to combine geological, geophysical, remote-sensing and drilling data into a 3D model of the subsurface, bringing resource and economic-value assessments forward with fewer drill holes; the endgame is to “use AI to make the Earth transparent,” then extend from technology services into mining rights, royalties and mining finance.
DeepOptica has moved beyond the pure-concept stage, but is “indeed still a little short” of the AI capabilities needed to support its full vision. The company completed a seed round led by Baidu Ventures in the second half of the year, has signed approximately RMB5.5M in orders and generated nearly RMB1M in revenue, with validation cases in Mongolia, the Middle East and Brazil; it plans to cover core mineral belts in Australia, South America and Africa this year, and expects to launch its first generalizable mining world model in the first half of next year.
The model’s key challenge is resolving the “non-uniqueness” between surface measurements and underground ore bodies—not layering a large language model onto the traditional workflow. 相子恒 wants to combine geological, geophysical and remote-sensing knowledge so that outputs conform both to local mineralization patterns and to actual physical measurements; roughly 50% of current training data is synthetic, and the next step is to expand the range of ore-body types covered by the synthetic-data engine.
The AI for mining startup referred to in the original as Cobalt, KoBold or Cobot has shown that AI can optimize drilling decisions; DeepOptica is trying to take the next step into asset valuation. 相子恒 summarizes that company’s output as determining where and at what angle the next hole has the highest probability of hitting ore, or where drilling would eliminate the most uncertainty; his view is that this still cannot answer how many tonnes of copper or ounces of gold lie underground, which is why DeepOptica wants to build a “world model for mining.”
The market entry point is not a handful of mining majors, but thousands of listed miners and junior miners without advanced technical capabilities. These small companies often have only a few employees and limited access to powerful AI systems or data solutions, while mining is not a monopolized industry; DeepOptica therefore plans to productize its services first, then, depending on regional understanding and project opportunities, take equity stakes through technology contributions, jointly acquire nearby mining rights or participate in mining operations.
The real moat is defined as data and a cross-functional team capable of building the entire stack from mining synthetic data to a mining world model—not the engineering ability to call open-source models. 相子恒 believes regional models require far less capital, while a true “Earth-scale foundation model” would demand enormous funding; the company wants AI to amplify a core team of 10-plus to 20 people and become a “small in scale but exceptionally cool, deeply rooted in exploration technology” technology-mining company.
相子恒 believes the company is most likely to fail on execution pace, not on a lack of ambition. The team aligns technology, commercialization and fundraising around three-year long-term, 18-month mid-term and three-month short-term plans; the method he brought from the Cambridge rowing team is to break goals down, lead without authority and build trust through action—“whether people like you as a leader, the correlation is not decisive.”
The five-year vision is to combine technology, mining operations and mining finance, while the early talent window is the scarcest lever today. 相子恒 envisions an “AlphaFold-level” open-source foundation model alongside a commercial version, then using its valuation system to invest in mining royalties and rights; 柯记’s warning is that a startup can use “hope” to attract people large companies cannot hire, but by years 3 and 4 that leverage “may disappear instantly.”

Deep dive

1. Mining Is Ultimately About Using Technology and Finance to Explore the Subsurface

相子恒 reframes the industry: “Musk is exploring space; mining is exploring the Earth—specifically, the world beneath the surface.” Leaders of international mining companies often come from technology or finance, not from the stereotypical image of the “mine boss.”
His industry view is that minerals are nonrenewable, while data centers, renewable energy and robots all require large quantities of metals, with human demand continuing to rise. Exploration is therefore not merely a resource business; it is upstream infrastructure for the energy transition and sustainable development.
DeepOptica’s one-line positioning is “an AI-driven company that aims to make the Earth transparent”: first, discover and assess economically valuable ore bodies faster and more accurately; over time, extend that capability to global resource intelligence.

2. Traditional Exploration Starts with Surface Clues but Ultimately Depends on Drilling

The first step, as 相子恒 describes it, is field reconnaissance and geological mapping by geologists looking for magmatic activity, mineralization events and mineral outcrops. An outcrop is underground mineralization brought to the surface by prolonged weathering, leaving a visible clue.
The next layer comes from geophysics and remote sensing. High concentrations of iron, copper and other elements can distort magnetic, gravity, electromagnetic or resistivity fields; alteration minerals leave spectral signatures that satellites can detect at the surface.
These signals provide only a hazy view of whether an area could host mineralization and what type it might be. Companies ultimately still have to drill to confirm the lithology and metal content at different underground depths. The show opens by characterizing mining as a business where a single decision can cost hundreds of millions of dollars—and where, once work begins, there may be no turning back.
The host’s question goes to the product boundary: if “making the entire Earth transparent” sounds almost insane, DeepOptica still has to begin with slices. 相子恒 acknowledges that the path starts with economically valuable ore bodies, but says global transparency is not merely a marketing metaphor.

3. AI’s Core Job Is to Collapse Countless Underground Possibilities into a Usable Structure

A single-modality model cannot absorb the full body of exploration knowledge. DeepOptica is trying to combine geology, magnetic fields, gravity, electromagnetics, resistivity, remote sensing and drilling data so the model understands what a given physical anomaly means under local geological conditions.
The difficulty is inverse-problem non-uniqueness: the same set of surface magnetic measurements may correspond to many different 3D underground ore bodies. The model must constrain the solution to a structure that both matches local mineralization characteristics and satisfies the physical quantities actually measured.
相子恒 divides the objective into two layers: first, identifying where mineralization may exist underground; second, determining the ore body’s 3D geometry, resource volume and potential economic value. “With fewer drill holes and more powerful data,” learning sooner than competitors how much copper and gold are underground is what changes mining decisions.

4. World Models and Synthetic Data Help Fill the Gap in Scarce Underground Ground Truth

The inspiration from world models is to incorporate rules and temporal processes into prediction. An ore body is not a static image: plate boundaries create conditions for magmatic activity; magma carrying high temperature and pressure moves toward the surface, reacts with surrounding rock and soil, and causes metal ions to exchange and form mineralization.
相子恒 believes these physical laws in Earth science and the way world models predict how objects evolve are “a fairly adjacent technology stack.” But he is explicit that existing general-purpose architectures differ substantially from a mining world model. The company will borrow selected small modules, while building the exploration model itself.
Roughly 50% of current training data is synthetic. The team compares known ore-body structures with synthesized 3D subsurface geological states to improve generalization; its synthetic engine may already perform well on typical volcanic-hosted mineralization and porphyry belts, but iron-oxide, skarn and sedimentary-hosted deposits still require sustained investment.

5. The AI for Mining Startup Validated Drill Selection; DeepOptica Wants to Value Ore Bodies

The industry benchmark cited by 相子恒 is an AI for mining startup, whose name appears in several forms in the original transcript. Its typical method matches existing drill holes with surface measurements, then uses Bayesian decision-making to determine the next hole’s location and angle.
The system produces two outputs: maximizing the probability of hitting ore, and eliminating the most uncertainty. Together, they can confirm an area’s mineralization potential at lower cost. The representative case cited by 相子恒 is the company’s discovery of a large, world-class copper deposit near an existing mining area in Zambia.
The host’s pushback is worth preserving: if this approach is already a generation ahead of traditional technology, why not simply replicate it? 相子恒’s answer is that decision intelligence is also on DeepOptica’s product roadmap, but “where to drill next” cannot tell a mining company how many tonnes of copper or ounces of gold lie underground.
The two companies therefore operate at different levels. The other company’s case centers on drilling decisions and mining-rights opportunities; DeepOptica wants to reveal a 3D structure usable for resource estimation and economic evaluation—“not simply giving a mining company the location of its next drill hole.”

6. “Earth CT” Is Harder Than Medical CT Because Observation Is Possible from Only One Side

相子恒 places exploration alongside protein and materials design as the same class of inverse problem: complex 3D systems in the real world typically leave behind 2D or n-dimensional measurements, and AI must infer the structure most likely to produce the observed properties.
Medical imaging offers the clearest analogy: 2D organ images must be matched to 3D organ structures. The team has also brought in members with medical backgrounds and uses ViT, multimodal feature extraction and related methods to infer the possible depth corresponding to a given signal.
But the host’s CT analogy remains incomplete. A person can be photographed from all sides; the Earth can only be measured at the surface. The solution is to expand the spatial range of surface sampling, trading horizontal information for insight into underground depth.
Language models and agents can process materials on resources, reserves and their corresponding value, taking on part of an analyst’s work. The ceiling for the exploration model is still determined by Earth-science data and physical constraints, not text-generation ability.

7. The Founding Team Answers Industry-Experience Doubts with “Understand the Need, Exploration and AI”

DeepOptica’s 3-person founding team crosses disciplines. 相子恒 has worked on remote-sensing satellite projects and served as a PI or technical lead on projects related to the European Space Agency and UK Space Agency. CTO 方博 previously worked with him on a remote-sensing startup and was an AI leader at a major UK financial institution. CFO Jesse was formerly a vice president at China Gold International and also has investment-banking and Greater China mining experience.
相子恒’s move from quantum optics into mining was not a complete leap. He studied quantum sensing, whose important applications include quantum gravimetry and quantum magnetometry; geophysics likewise deals with gravity, magnetic fields and partial differential equations, with analytical methods that overlap with quantum physics and wave optics.
The most common fundraising objection is that Jesse has more than a decade of mining experience, while 相子恒 and 方博 together have only about 5 or 6 years of industry exposure. Why should they be the ones to remake a traditional industry? His answer is not to deny the gap, but to have Jesse define what users need, while he owns exploration technology and 方博 owns AI.
相子恒 acknowledges that this is not the only viable team structure, but believes a geologist with 50 years of experience may be constrained by inherited frameworks, while an outside perspective can produce new data-driven solutions. An interesting parallel is that the other AI team’s profile also combines quantum computing, Earth science and finance.

8. Small POCs Are the Wedge into Overseas Enterprise Customers

The company completed a seed round led by Baidu Ventures in the second half of the year, signed approximately RMB5.5M in orders and generated nearly RMB1M in revenue. Validation projects and commercial orders are already spread across Mongolia, the Middle East and Brazil, with customers typically starting through proof-of-concept projects to test the model and data-processing capabilities.
相子恒 says Brazil’s G1-21 is the largest geological consultancy in South America and one of DeepOptica’s largest customers; the two sides are already collaborating at both the data and product levels. For a Chinese team at the seed stage, the relationship itself is evidence of early technical credibility.
Overseas trust starts with delivery. After a Middle Eastern customer raised a need, the team quickly completed an initial analysis, prompting the project manager to tell the partner: “See, I told you, this team can manage a lot of stuff.” The baseline 相子恒 emphasizes first is simply being reliable.
His cross-cultural approach is not to rush to prove himself, but to understand within a minute who the other person is, what they want and why they are willing to talk, then present a solution through a warm introduction. “First create a heart-to-heart connection”; only then does the actual business become smooth. He also believes AI’s development has made people realize that “Chinese people in AI are the most reliable.”

9. Junior Miners Provide the Product Entry Point; Mining Rights and Royalties Offer Long-Term Upside

Asked whether this is a crowded field or a blue ocean, 相子恒 gives a dual answer: internally, the team pressures itself as though it were operating in a crowded market; from a broader perspective, large numbers of problems and customers globally remain unserved.
There are thousands of listed mining companies worldwide, and many junior miners have only a few employees, putting powerful AI systems and data solutions well beyond reach. Large enterprises are developing more efficient exploration methods, but mining is not monopolized, leaving room for productized solutions serving a fragmented customer base.
The host frames the strategic fork as “sell software” or “become mine owners ourselves.” 相子恒’s sequence is to build the model and product first. Once the team understands a region, it can continue serving customers or jointly acquire nearby mining rights and take technology equity in mining companies.
This conversion does not require waiting for a globally generalizable model. If a high-potential target area appears around a customer’s existing mine, DeepOptica could use project data to “claim the hilltop,” partner to acquire nearby mining rights and turn its technical advantage into an asset stake.

10. Regional Models Can Advance Leanly; an Earth-Scale Model Still Requires Heavy Capital

The company’s short-term North Star is not model scale, but 2 things: under what conditions the technology solves a customer’s problem and with what accuracy, and whether the product truly resonates with customers. 相子恒 is relatively confident in commercialization because once the model reaches a sufficient capability threshold, it can serve many regions and mining companies of different sizes.
The roadmap is to complete validation in core mineral belts across Australia, South America and Africa this year, then launch the first generalized mining world model in the first half of next year. Mining-rights partnerships will proceed according to project opportunities and do not require waiting for strong model generalization.
Asked whether the technology will eventually become a commodity, 相子恒 says that will be “quite difficult.” Capital and talent will limit participation; the deeper gap is that Earth science and computer science have not yet formed a mature interdisciplinary system comparable to physics-computing or biology-computing. A few engineers and a small amount of data from one mining company are not enough.
He distinguishes between 2 levels of capital intensity: regional generalization requires far less investment, while a true Earth-scale model will “certainly require enormous investment.” The ambition remains to build a DeepMind-like company applying the technology to mining, rather than stopping at project-based consulting.

11. The Small-Team Thesis Depends on AI Amplifying Productivity, Not Avoiding Fieldwork

相子恒’s ideal company is “small in scale but exceptionally cool,” with a core team of perhaps 10-plus to 20 people; overall, he believes a team of 20 or 30 may already be enough. The point is not headcount, but finding the people who matter most, using AI to maximize productivity and letting members enjoy what they do every day.
The host’s objection is practical: large customers usually require one-by-one service, while mining appears inherently labor-intensive. 相子恒 points to the large Middle Eastern service order as a response. With AI-driven productivity gains, a single project does not require the traditional headcount; when fieldwork is necessary, the team sends people to complete clearly defined tasks, then brings the data back for centralized analysis.
His conclusion is that this working model could replace part of traditional consulting services. The team is still hiring across Earth science, AI synthetic data and foundation models; a lean organization does not mean ignoring domain experts.

12. Unmandated Leadership, Execution Pace and the Talent Window Will Determine the Endgame

At Cambridge, 相子恒 spent 4 years rising from the men’s second boat to the first boat, then becoming vice captain, college rowing president and club captain. His personal performance ranked around 2nd or 3rd in the college, but he developed a deeper understanding of training organization and team improvement. He describes the style as “a very calm form of leadership.”
Rowing imposes extremely concrete management constraints: during his PhD, the team trained twice a day, and if any one person was absent the entire boat could not launch. In the spring, teammates confirmed at 5 a.m. in the group chat; anyone who failed to respond got a knock at the door. The team assembled at 5:30 a.m., aiming to be the first boat on the river. Goals were set before training, each person’s 2,000-meter improvement was broken down month by month, and everyone ate together afterward.
His identity as a Chinese captain and his less-than-dominant physical performance both drew skepticism. He established legitimacy by securing better coaches and equipment, transporting the boat roughly 100 km to compete in a race they otherwise would not have entered, and ultimately winning a Cambridge regatta. Faced with a Russian teammate who disliked him, he treated that person simply as “someone I needed to help improve.” “Whether people like you as a leader, the correlation is not decisive.”
Asked about DeepOptica’s biggest risk, his answer is “very likely the pace”: fundraising, customers and technical milestones all have to land within their windows. The team uses a 3-year plan to govern an 18-month plan, then breaks it into executable 3-month targets. Responding to the host’s joke about being a “small-town exam grinder,” he insists the company is more like a “well-organized vehicle” driven by long-term vision.
The 5-year success picture is an “AlphaFold-level” mining model with both a foundational open-source version and a commercial version, working deeply with several of the world’s largest mining companies. The company would then use the model and its valuation system to invest in mining royalties and take stakes in mining rights, combining technology, mining operations and mining finance. “I think we will definitely achieve it within 5 years.”
That confidence does not come from an uninterrupted run of success. 相子恒 recalls setbacks in his PhD experiments, the frustration of peers producing more influential research, and a 2021 startup that was “not successful at all.” Failure did not make him more conservative, because his anxiety was not about having less money or a stagnant standard of living, but about “not doing what I want to do.”
His longer-term ambition remains to explore both Earth and space: from remote sensing and low-Earth-orbit satellite propulsion to mining, then onward to space mining, seabed exploration and “even crazier things.” 相子恒 identifies resilience, leadership and unmandated leadership as key traits for early-stage founders.
柯记’s closing advice is explicitly time-sensitive: startups can use “hope” to recruit people whom large companies struggle to hire, but after years 3 and 4 that leverage may disappear through failure or scaling. The window therefore calls for being “extremely ambitious” early, even reaching out to people who appear impossible to recruit; they may be in a period of career anxiety and willing to try once.

Verification Notes

The original uses several names for the cited AI for mining startup—“Cobalt Metals,” “Cobot Metal” and “KoBold Metals”—and the transcript alone does not establish which is correct. The main text therefore refers to it as “the company.”

Monday.com CEO on Is SaaS Dead: Will Everything Be Vibe Coded | Eran Zinman Brian Armstrong: When Washington Tried to Kill Coinbase