top of page
Search

When the Smartest AI Model Is Worth the Least


At SuperAI 2026, a European lab, a Chinese lab, and a veteran investor agreed the AI frontier has quietly moved — away from topping benchmarks, toward owning deployment, distribution, and the ground beneath the tokens.


A panel discussion, moderated by Zixuan Li (Head, Z.ai), with:

Geoff Soon, Mistral AI  ·  Cherie Shi, MiniMax  ·  Hemant Mohapatra, Lightspeed

SuperAI 2026  ·  Marina Bay Sands, Singapore  ·  10 June 2026


For a few years, the word frontier had a simple meaning in artificial intelligence: whoever sat on top of the benchmark leaderboards that week was at it. The position was expensive to hold, it changed hands constantly, and an entire industry organised its marketing around the chase.

That definition is now coming apart — and the people dismantling it are the ones who build and fund the models.

Across forty minutes on a Singapore stage, a sales chief from Europe's flagship lab, a manager from one of China's fastest-rising labs, and a partner at a venture firm that has watched the whole cycle arrived at a shared and slightly vertiginous conclusion. Raw intelligence is on its way to becoming a commodity. The smartest AI model in the world may soon be one of the least valuable places to stand.

The panel — “The Global Frontier of AI Models” — was moderated by Zixuan Li, who heads Z.ai. The voices were Geoff Soon, who runs revenue across Asia-Pacific for Mistral AI; Cherie Shi, a global business manager at MiniMax; and Hemant Mohapatra, a partner at Lightspeed who supplied the conversation's connective tissue — a set of frameworks about where value goes when a new resource stops being scarce.*

The frontier stopped being a leaderboard

Asked what “frontier” now means, none of the three reached for a benchmark. Soon argued the logic had inverted over the past year. Topping a leaderboard is costly and fleeting; what his customers actually care about is frontier deployment — what a government or an enterprise can put into production to drive real change. And in a world of chip shortages and uncertain API access, that question is inseparable from ownership: how much of the capability you own, how much you control, and how durable it is if the supply chain takes a shock.

Figure 1: Frontier' has quietly shifted from reading a chart to owning what you can actually deploy (Source Wikimedia Commons).

Shi's reframing was about to reach. For MiniMax, she said, the frontier is measured by how many people in the real world use the models every day — which is why the company optimises relentlessly for two things: agentic capability on genuine day-to-day tasks, and cost efficiency low enough that a user in any corner of the world can afford frontier-grade intelligence. The benchmark she cares about is adoption, not the leaderboard.

Mohapatra zoomed out furthest, declining to answer directly and instead handing the room a lens. Every technological super-cycle, he said, opens with an extractive phase: you are pulling a new natural resource out of the ground, and decades of research go into finding it. Three hundred years ago, that resource was oil. Today it is intelligence — the AI token. In the extractive phase, the frontier is wherever you sit closest to the extraction, because that is where the value pools.


Intelligence is the new oil

The trouble — or the opportunity — is that extractive phases end. As the resource gets cheap and abundant, the cycle turns distributive.

You stop competing to pump the most oil and start competing to put oil into machines and build better cars.

The frontier migrates with it: being the best token-extractor stops mattering, and the question becomes whether you can actually solve a customer's problem. By Mohapatra's reading, that turn is happening now.

Figure 2: Every super-cycle opens by extracting a new resource. Photo: Lakeview No 1 Gusher 1910 (Source: Wikimedia Commons)

Watch the model layer commoditise in real time, he suggested: a model released the day before the panel was among the strongest ever, yet the gap between one generation and the next keeps shrinking. That is inevitable as raw intelligence approaches a ceiling, where each additional one percent costs ten billion dollars, then a hundred billion.

Past that point, models compete on price rather than capability — and the only players who can win a price war are the ones with the balance sheets to fund it: the big clouds, OpenAI, xAI, Anthropic, Google, Microsoft.

They will drag the model layer down to commodity economics. And, he stressed, you want that to happen, because it pushes the value up the stack.

Commoditisation doesn’t mean cheap or free. It means fungible — oil from one station is no different from another.  — Hemant Mohapatra, Lightspeed

That distinction did a lot of work. A commodity is not worthless; it is interchangeable. Intelligence from one model becomes indistinguishable from intelligence from another, and the competition collapses to price against performance — exactly as it did for oil. Along the way, Mohapatra offered a deliberately provocative yardstick for artificial general intelligence: in his view, an AGI is something like an IQ of 98, a system that makes roughly the mistakes a human makes but carries human intuition. You then take that average-human mind and train it up to 150 on mathematics, on biology, on agentic work — until it becomes a specialist. Generality is the floor; the value is in what you build on top of it.


The money is in scope, not scale - Smartest AI Model

If the baseline models will soon solve the easy problems for everyone, where does a new company find an edge? Here Mohapatra drew his second line, between two kinds of problems. A scale problem is one you solve once and then deliver cheaply and abundantly to many — coding is the canonical example. A scope problem is messier and keeps moving: cancer research, mathematics, biology, material science. You can train an intelligence to beat one cancer, and then a new variant appears and your solution fails, so you iterate again, forever.

Figure 3: Scope problems — biology, material science, mathematics — keep moving, and that's where the value migrates Photo: Molecular model of Penicillin by Dorothy Hodgkin. (Source: Wikimedia Commons)

Baseline models will own the scale problems by default. So the firms worth backing now, he said, are the ones climbing a different intelligence curve entirely — physical world-models, large biology, material science, mathematics, physics, game simulation — fields where the token output looks nothing like a chatbot's. That is where the value has gone.


The Swiss Army knife and the specialist tool

Soon translated the same idea into a tool metaphor. A giant general-purpose model is a Swiss Army knife: perfect when you have a rough idea and want to prove out a hypothesis, flexible enough to get you there even if it is not the most efficient route. Once the use case is proven, you reach for a specialist tool — a smaller, sharper model that does one thing efficiently, which is where labs like Mistral and MiniMax come in. But the level that matters most, he added, is the one above the tool: the user, and the enterprise context you can feed into the specialised model to produce a real outcome.

Figure 4: A giant general-purpose model is a Swiss Army knife (Source: Wikimedia Commons)

This is also why Soon is sceptical of a single, monolithic super-intelligence. He expects many domain-specific super-intelligences instead. Within twelve to eighteen months, he predicted, AI may well perform many functions of a traditional call centre at superhuman levels — while a system capable of diagnostic surgery remains far off, with too many unsolved steps in between. The competitive question for a smaller lab is therefore not how to out-muscle a general model, but how to specialise knowledge, run it efficiently, and wrap enterprise context around it so a customer builds a genuine moat instead of merely renting a generic intelligence for a 10-to-30-per-cent productivity bump.*


Two labs, two playbooks

The labs on stage are pursuing that scope-and-specialisation frontier from opposite ends of the world, with strikingly different advantages.


Figure 5: Two labs, two playbooks

Mistral: Europe’s third option

Soon leaned into geography. Europe is dense with industrial heavyweights, and partnering with them lets Mistral push to the frontier of fields like material science. At its first customer summit in Paris, the company announced a strategic partnership with Airbus — the kind of partner sitting on vast proprietary data that can be fused with Mistral's models to attack genuine scope problems. He was candid about the disadvantages too: a real gap in funding and liquidity versus a Silicon Valley that has refined capital-raising for decades, and a home market that is effectively 27 countries trying to behave as one, against the single regulatory and linguistic blocks that China and the United States enjoy. His closing point was almost diplomatic — a reminder that the world is not a two-horse race between China and the US, and that Europe is a credible third option for anyone choosing an AI partner.


MiniMax: the cheapest intelligence in the world

Shi's differentiators were modality and price. MiniMax built multimodal from day one — not just language but Hailuo, its video-generation model, plus speech and music — and she sees those modalities converging into single models that improve both generation and understanding at once, something she said very few companies can do. Its recent M3 release is pitched as both a strong agentic, tool-using model and a natively multimodal one. The second pillar is cost: she claimed M3 runs at roughly one-third to one-fifth the price of Sonnet 4.6 while handling comparable coding and agentic work, part of a deliberate strategy to keep every release among the most cost-competitive in the world.

Behind the pricing sits an argument about return on investment. Many enterprises and AI-native users, Shi noted, are enthusiastic about models but have not yet had the hard internal conversation about ROI. When they do — and she believes they will — MiniMax wants the math to still work even at enormous token volumes. The company runs a hybrid model strategy to match: its language models, from M2 through M3, are open source, while video, speech, and music stay closed, where customers still look for a defining frontier. Open weights, she explained, are what let MiniMax serve the booming demand across Asia-Pacific and Southeast Asia for models deployed in local data centers under strict privacy rules — though for the very newest capabilities, the cloud version always ships first.


Can you make money giving it away?

That hybrid strategy opens onto the sharpest disagreement-by-degrees of the session: the economics of open source. Mohapatra split it cleanly into a philosophical case and a business one. Philosophically, you want AI available to everyone and the foundational layers commoditised rather than controlled by a handful of corporations charging the maximum, and open weights are simply the fastest route to that commoditization, because they let everyone tune and shape the intelligence. Closed models get there too, eventually; oil is controlled by a few players and is still a commodity. Open source just speeds the clock.

The business case is where it gets hard. People want to consume intelligence on tap — open the tap, fill the glass — and the only way to make money selling tokens that way is to own what is under the tap.

You can’t just sell water from a pond. You have to go all the way down to the GPU, the energy, the land.  — Hemant Mohapatra, Lightspeed

For a company selling an open-weight model as a model, that is a trap. Today open source is, for some, mostly a marketing posture — the moment there is something worth protecting, it goes private. The state-of-the-art gap is closing fast: six months ago open models trailed the closed frontier by six months, now by three, and the lead will keep shrinking until nobody pays a premium for a three-month edge that costs ten times as much. Which leaves the uncomfortable question Mohapatra does not think has a happy answer: can you make money in open source without owning every layer beneath it? He does not believe you can. Soon, notably, agreed — the way to monetise this technology, he said, is to win the full stack, from the GPUs through the inference layer to the harnesses on top.

Ford, GM, and the horses

To explain what kind of team wins as a cycle matures, Mohapatra went to Detroit. Henry Ford was an extractive-phase champion — any colour you like, as long as it's black — whose genius was scaling a single product flawlessly.

Then General Motors arrived in the distributive phase with a different proposition: not whether you owned a car, but whether you owned one that stood out, a V6 against a V4. GM's real innovation was closer to planned obsolescence — making last year's styles feel dated and rolling out new ones — and it was a market-and-brand company far more than an R&D one. The lesson he draws is that the winning playbook changes with the phase, so he evaluates founders on two axes at once: where the cycle is, and whether the team fits the cycle's moment.

Figure 6: Unidentified Location: Queensland, Australia Description: Ford Model T, one ton capacity truck (Source : State Library of Queensland)

Which set up the panel's closing round — each speaker's most underestimated success factor, in a sentence. Soon's was distribution: having watched the scramble for GPUs, he thinks access to this capability could be just as disrupted in the years ahead. Shi's was a mission more than a factor — to distribute frontier intelligence to every user in every corner of the world by driving the cost of tokens, energy, and infrastructure relentlessly down. Mohapatra's was the most pointed and the best note to end on.

This resource, he said, has simply been handed to us; things that took weeks now take minutes, and tasks that took twenty steps can be done in one. Yet far too many founders are taking this miraculous new substance and using it to make the old thing marginally faster. Even companies that race from one to three million dollars in revenue in a year — a feat that was nearly impossible not long ago — are being passed over, because others are moving faster still, with better retention and a bigger vision. The resource is not the differentiator anymore. The imagination is.

Too many founders take this amazing vial of oil and feed it back to the horses, hoping they’ll run faster. They won’t.  — Hemant Mohapatra, Lightspeed

The Samso Way – Seek the Research

Here at Samso, we pride ourselves on delivering content for investors that is independent and informed by over three decades of experience in the industry. Our content is well-researched and is only created if I see merit in discussing the company's story.

Our mission is simple: cut through the noise and spotlight what matters—genuine stories, grounded insights, and real opportunity.

Our content is well-researched and is only created if the team sees merit in discussing the company or concept. Investors can explore our three core platforms: 

There may be numerous paths to success in investing, but the common thread among successful individuals is that they remain committed to making informed decisions. Equip yourself with the right knowledge and tools, and you will be well on your way to achieving your financial goals.

Most importantly, investors need to be absolutely diligent in understanding their own risk-reward tolerance and capabilities. Never bite off more than you can chew. As they say, Rome wasn’t built in a day, and the Great Wall stood because it took centuries to complete.

The Samso Philosophy:

Stay curious. Stay sharp. And remember—digging deeper always uncovers the real value.

In Life, there is no such thing as a Free Lunch.

Never bite off more than you can chew is my parting comment.

Happy Investing, and the only four-letter word you need to know is DYOR. 

To support our independent nature of our work, please head over to our Support Page and give us a helping hand in any of the ways listed. This is a new initiate for the Samso Platform, and it was always the concept of Samso when we started this journey in 2018.

Disclaimer

The information or opinions provided herein do not constitute investment advice, an offer, or solicitation to subscribe for, purchase, or sell the investment product(s) mentioned herein. It does not take into consideration, nor have any regard to your specific investment objectives, financial situation, risk profile, tax position and particular, or unique needs and constraints.

Share to Grow: Your Bonus

Samso has just released an eBook: How to Add Value to your Share Portfolio


Download eBook | Samso Insights
Download eBook

If you find this article informative and useful, please help me share the information. I try to write about topics that are interesting and have the potential to be of investment value. It is not easy to find stories that fit those parameters. If you or your organisation sees the benefit of what Samso is trying to achieve and has a need to share your journey, please contact me at noel.ong@samso.com.au.

Samso is a trusted platform that equips dedicated investors with up-to-date industry knowledge and insigh0ts from top CEOs and thought leaders. By staying informed on business advancements and market trends, investors can enhance their financial decisions through a combination of expert guidance and their own research.

Samso News  |  www.samso.com.au  |  An Investor Lens on ASX-Listed Companies

Comments

bottom of page