Decentralization has always been a cornerstone promise of crypto. Today, as we near Bitcoin’s 20-year mark, it seems the very asset class that promised decentralized finance has become over-centralized instead, and the only parts left of decentralization are among crypto’s weakest points.

Somehow, crypto ended up with several centers and seems even more centralized than the system it was meant to replace. Today, the crypto industry depends heavily on a few players, with decision-making, value, and control highly centralized.

Decision-making is centralized around foundations, whether these are the organizations responsible for the development and advocacy of different cryptocurrencies and blockchain projects or for-profit companies issuing and managing these projects.

In both cases, these foundations are usually the largest holders of the cryptocurrency. Although most permissionless blockchains offer a governance mechanism that allows voting these organizations out of control, we still haven’t seen this happen. Attempts to revoke a foundation’s proposed changes usually result in a less popular, less valuable chain split.

Ethereum Classic (ETC) was born when a part of the Ethereum community rejected the reversal of the transactions of the DAO hack, and Ethereum PoW (ETHW) was forked by the community that rejected the shift from PoW to PoS consensus mechanism. Bitcoin’s community disagreements resulted in Bitcoin Cash, Bitcoin Gold, Bitcoin SV (short for Satoshi Vision), and a long list of chain splits (sometimes from other chain splits).

Long story short, decentralized decision-making rarely works against the will of the foundation behind the blockchain, and while this results in unpopular changes in the case of most major layer-1 chains, it results in less agility and ability to update or upgrade where the foundations are weak, such as in the case of Bitcoin, where most chain splits happened over wanting, rather than rejecting, changes.

The result is the same: a non-dividing initiative can only be taken by a central organization.

Control and custody are centralized around exchanges. Well, two exchanges in particular. Binance and Coinbase host up to $504B worth of crypto assets, according to DeFiLlama and SEC numbers (Coinbase self-reported). That’s the equivalent of almost 20% of the total crypto market cap as calculated by CoinGecko. When the news about Binance’s hack was released, Bitcoin saw a drop of over 4% within hours.

When FTX, the second-largest crypto exchange then, collapsed, Bitcoin, which was already around 70% away from its one-year-old all-time high, lost 26% of its value in almost 3 days; by default, it dragged the rest of the space with it, which I’ll discuss in a bit.

This exchange centrality stems from a simple fact: the tech needed for decentralization to work is so complex for regular users. A mediator is needed, and just like crypto’s nemesis, banks, this mediator needs to have some sort of control over your money.

Value is centered around Bitcoin. Although blockchain technology saw some significant developments over the last two decades, since May 2021, Bitcoin’s dominance has never gone below 40%. Out of ~$2.6T in crypto market cap, Bitcoin’s share is $1.6T. In other words, Bitcoin’s value is larger than all the other crypto projects combined by over $400 billion, including stablecoins, wrapped coins, and any shitcoins that made it to the list of over 22,000 projects observed by CoinGecko.

And even this one asset, overshadowing the whole space, is quite concentrated in terms of ownership. The top 100 identifiable owners of Bitcoin hold around 9.18% of the max supply of 21M BTC. If we add Satoshi Nakamoto’s estimated holdings, this number rises to over 14%. Yes, 1 of every 20 bitcoins is owned by one anonymous person.

Bitcoin’s centrality means other projects’ attraction of more investment and better chart performance is dependent on an external factor. Ambitious blockchain projects like Ethereum, Polkadot, or Solana will always see their largest upsides and downsides as side effects of Bitcoin’s movements, which affect, rather than reflect, the general sentiment on overall crypto.

This level of concentration tells us that crypto is not a “space” where projects are competing. It’s rather somewhere between a single asset and an asset class, mostly closer to the former end of the spectrum.

The only part left from decentralization is the worst: transaction irreversibility, which makes crypto hacks and thefts much worse and harder to mitigate. While blockchains are still immune from attacks so far, most crypto adopters rarely use the blockchain directly, but rather use intermediaries, all of which have been hacked one way or another.

Self-custody software gives the illusion of control, but it still can be—and has been— hacked. More than 5,000 Atomic wallets were compromised in 2023, with over $100M worth of crypto stolen. Trust Wallet suffered vulnerabilities in 2023 that led to stealing $170k worth of crypto, and again got attacked in 2025, with estimated losses of around $7M.

However, decentralization is not the only promise that crypto failed to deliver on. Two other promises were broken: privacy and defeating traditional finance.

In exchange for hopes of more moonshots and making astronomical profits, crypto longed for more regulation and integration within the traditional financial systems. The extensive KYC process any user has to go through today to buy crypto safely not only eradicates the privacy features, but, ironically, makes traditional finance provide more privacy than DeFi. Your bank won’t be publishing a record of your transactions on a publicly accessible ledger.

As for the promise of bringing global finance to its knees, well. Both Wall Street and crypto bros shout in anger at the interest rates lately, with ETF net inflows becoming more of an optimism factor than public adoption of crypto. This ship had sailed a while ago already.

So… (uncomfortable silence) what’s left of crypto?

It’s the least decentralized projects, such as Circle, Tether, and Ripple, that help freeze and recover stolen assets. Crypto projects with some degree of centralization have been on the rise in the last few years.

If we exclude Bitcoin’s market cap, cryptocurrencies with some kind of centralization comprise about 52% of the top 100 crypto projects’ market cap. Stablecoins alone comprise over 26%, with a ~$280B market cap.

The trend is kind of clear, especially if we count government-run central bank digital currencies (CBDCs) that have started to move from headlines to actual implementation. According to the Atlantic Council’s CBDC tracker, 146 countries are exploring or involved in CBDC projects, 41 of which are in the pilot phase, including China’s massive e-CNY, which has processed 3.4 billion transactions in 2025, worth a whopping $2.3 trillion.

While crypto failed to create the financial anarchy it once hoped for, it seems like it ended up enhancing the financial system it aimed at. The advantages of security, speed, and lower transaction costs it brought are only going to go further, with much of its contributions happening behind the scenes, providing institutions with better, more cost-effective channels for record-keeping, transfers of value, and settlements.

Pushing for the crypto revolutionary dream is a lost cause. However, pushing institutions to pass down the benefits that crypto and blockchain technology brought, instead of using them as a profit margin boost, is not.

Do you get advice from AI? We need to talk.

AI is good at finding answers, and it’s just going to keep getting better. Even if the pace of improvement slows down, we’re very unlikely to deal with a worse AI soon, unless regulation and safety measures become so extreme they move the needle backwards.

However, answers we get from AI aren’t all the same, and identifying which type you’re looking for determines how you treat the answer.

There are mainly two types: The first is information, and the second is consultancy.

When asking for the right size of a screw, the cheapest music streaming app subscription, or how to disable video autoplay on Facebook, the answer is information.

AI is great at finding information online and organizing it, saving us the time of juggling several websites, making comparisons, and connecting the dots ourselves. Since web search became an integrated skill of each frontier model, cutoff times and hallucinations became much less of a problem.

AI still can misinterpret some sources, and it’s always a must to check the sources yourself, but generally, AI apps have become more reliable in this regard.

On the other hand, when asking for the best path for your business or gift recommendations, the answer is consultancy or advice.

The difference lies simply in whether you’re asking for a fact that can be proven right or wrong, or asking for an opinion or an analysis. In other words, would the core of the answer to this question be the same from different AI models and humans or not?

Advice, consultancy, or ‘thoughts’ that come from AI aren’t necessarily bad. In fact, in many cases, they’re just stating obvious best practices or basics people tend to overlook. However, understanding the inherent flaws in the data on which this advice is built and how it’s interpreted and utilized helps us have a healthier perception. Some of these flaws are data-related, many of which are problematic regardless of AI.

Let’s break down the process an AI model goes through to provide a consultancy answer into three stages, and highlight the flaws that might affect each of these stages.

Information & data

↓

Interpretation & conclusions

↓

Recommendations & consultancy

These three stages don’t necessarily happen in this order, and sometimes two of them happen at the same time, but for the sake of simplification, let’s assume the process is as simple as that, and look at the flaw threatening the integrity of each one.

Stage 1 – Data collection: picture imperfect

This stage doesn’t start with your prompt. It includes the data that an AI model was provided with for training, the data the model collects on the go via web search and other connections, and, of course, the data you provide as references or contexts.

While AI has been trained on tons of data, it’s worth noting an obvious fact: AI only has access to digital data. In other words, the only reality AI knows is that with a digital representation, and what gets to be digitized is neither random nor neutral.

For example, if we look at economic data, poor countries tend to have wider shadow economies, and thus a smaller digital footprint of their economic activities. So when AI is trying to understand economic concepts, it has more data on rich countries, and the little data it has on poor countries doesn’t only lack coverage, but lacks representation of reality altogether.

On an individual level, the same bias applies: people rarely post detailed case studies on how their business failed, or how they ended up in debt and burnout.

In addition to underrepresenting reality, many datasets suffer a well-known, systematic flaw: survivorship bias. This is when the data collection or sampling method leads to surveying only one type of cases.

A simple example is looking for successful businessmen who dropped out of school and using the results of this search to evaluate the decision of dropping out. The results would probably state that dropping out is a great decision; I mean… look at all those great billionaires in their black T-shirts. The problem with this data set is that it ignores all the ones that didn’t make it.

Stage 2 - Interpretation: ghost of data past

When interpreting the data that came from the previous stage, another flaw is possible.

Much of the evidence used in many fields where a ruler doesn’t help is built on observations of data and drawing correlations, rather than actual confirmed detection of causation.

To understand this, let’s visit the concept of A/B testing. This is a common optimization tactic in marketing that is very similar to scientific research: You take two inputs, keep every single aspect of both identical, and change only the factor you’re trying to study or measure its effect.

However, in marketing, many glorify “best practices” guides that different advertising and social media networks release. These weren’t built using the rigorous process above. The wording of these guides, or at least the research they cite, is not misleading.

For example, these guides would say there is a correlation between shorter content and higher engagement, but we don’t know that this causes that. Both could be results of a different cause, like brand size, which means people are more likely to engage with it, and at the same time it has enough resources to hire better writers able to express campaign messaging in fewer words.

Not to mention that these ‘bibles’ of advertising are mostly built on English campaigns, not taking into consideration, for example, the confusion fewer words could cause in some languages.

If you ask an AI to give you a work plan for a blog, it’ll build one on the patterns it spots across existing, successful, indexed blogs. The plan will include posting frequency, social media channel matrix, clusters of articles to rank for SEO, etc. However, this plan is just a description of the average of what similar blogs are doing, not what they’re doing to succeed.

Also, AI—and humans, to be fair—don’t draw correlation or causation from their own analysis all the time, but rather by copying others’ analyses or descriptions of correlations and causation, falling into the same flaws of the first stage described above. So theoretically, AI wouldn’t be giving you bad advice, just copying someone else’s bad advice with higher confidence.

Stage 3 - Recommendation: you’re a genius!

This stage is when the never-ending answer pops up, whether it comes as a fancy-designed deck or just text full of bullet points and tables. Usually looks like a well-knit, solid plan, although it might be complete rubbish.

Three main issues might make the plan you’ve received not as useful—or elegant—as it seems.

The first one is context blindness. AI can give you a solid marketing plan for your business but miss what your business actually does, either because you didn’t mention it or because the context window got so jammed that it missed that token.

For example, a minimal design would work for an Apple campaign, but would it work for a Disneyland campaign? Even if we want to look at similar brands, a Nike campaign element won’t necessarily work for Adidas. Something funny by one brand can be cringe by another. A saving promise is seen as a good deal by one brand, but a source of suspicion by the other.

These comparisons get trickier when your business or project is more niche, and you can rarely find similar projects with many elements in common.

This is not only an AI flaw, but it could be you. While asking for advice to deal with your flu or insomnia, you might miss mentioning some condition that sounds irrelevant to you, but for AI, or a real doctor, it would change the whole context.

Second is false precision; this is especially relevant when AI’s response includes numbers or probabilities. While sourcing numbers from other sources is a legitimate way to get more accurate answers, numbers the model itself calculates may provide a false sense of accuracy.

The third and most critical flaw at this stage is sycophancy (fancy word for sucking up). AI models are commercial products, after all, and customer retention is a contributor to how they develop and evolve. Nothing satisfies a customer like agreeing with them, especially on their most unreasonable ideas they can’t find humans to affirm.

Sycophancy might extend and compound beyond one chat, flow, or task. Frontier models now keep a memory of past conversations, which is a very useful feature most of the time. However, this might mean that sycophancy kicks in even if your ideas that an AI is agreeing with are mentioned in the same chat or flow.

So, what to do? How to benefit from the great capabilities of AI without falling into these traps?

Simply ask yourself: is the answer I’m looking for information or consultancy? If it’s consultancy, use the answer as a brainstorming process or starting point, rather than a final output.

A plan you get from AI is a plan in which you might invest your and others’ effort, time, and money. And like any investment, you should do your own research.

We stand at a weird point right now. AI’s contribution to everything online has gone as far as it can in almost 5 years of widespread use, while legislative and platform regulation hasn’t caught up; neither has our understanding of the hybrid digital reality we’re dealing with.

What we have online now is not a digital twin of humanity. If an alien civilization tries to understand us by looking at our content and systems available online, it will face a digital twin of a complicated blend of human, AI, and hybrid components.

While watermarks somehow help filter out some AI-generated content, the same doesn’t apply to a wider and more influential AI contribution: decisions.

For example, when a user is verifying their identity and address to complete the KYC requirements of a trading broker or a crypto exchange, it’s safe to say that the first round of review of the submitted documents is almost certainly performed by an AI. Single-minute verifications on some platforms suggest that this is the only round. That’s why sometimes we need to upload a different picture or scan that an automated system can read, even if it’s already easy to read by a pair of human eyes.

In this sense, and even though the company approving the account verification still bears responsibility, it’s not clear whether this approval is a human, AI, or hybrid decision.

The same applies to a faster and much more frequent activity: trading.

A thousand bids in an order book don’t necessarily mean a thousand traders willing to buy, but rather a thousand decision processes whose outcome is to buy, whether it’s a newbie’s trust-me-bro moment or an AI agent’s analysis of thousands of data points, including the order book itself.

In his great book Other People’s Money, John Kay provides a great analogy. He starts with a real 1906 experiment that led to the concept of the wisdom of the crowd, where people have to accurately estimate an ox’s weight. Kay extends this to imagine that the scale breaks and the official weight is to be decided by the average of everyone’s guesses, which was actually very close to the truth in the original experiment.

In such a scenario, people would stop trying to guess what the ox’s weight is, but rather what others would guess. This, pretty much, sums up financial markets’ main drivers today, especially that they, in fact, lack a neutral scale to weigh the real value of any given asset.

This leads to sentiment playing a big role beyond a clear, coherent equation. P/E ratio incoherence among the largest companies in the world is a good visualization of this chaos. P/E ratio of Nvidia is 28.8, Apple 38.5, Alphabet 17.3, and Tesla 324!

When traders, especially day and swing traders, place their orders, they’re not estimating the fair logical value of a share or an ounce of gold, but rather trying to get ahead of what they think other market participants’ estimates would be, or even what other participants’ estimates of other participants’ estimates would be.

Identifying support and resistance levels is about identifying other participants’ perceptions rather than an asset’s logical value range. The drop after an interest rate decision is mainly driven by expectations of others’ reactions, rather than actual investor reallocation. Examples are endless.

Today, market participants are legally human, but in reality they’re a mix of humans, algorithms, and AI.

High-frequency trading algorithms seem to be on the rise with the advent of AI and how much easier it has made it to build, for instance, an MT5/4 Expert Advisor. Once an institutional advantage, sophisticated trading automation is now a normal tool retail investors and traders can have.

In addition to (ironically now considered) simple coded algorithmic trading, we now have more sophisticated multi-step agents. These are not only reliant on quantitative input, but also parse news, social media content, and other qualitative sources.

Glitching glitches

Anyone who has been long enough in the market knows that the markets sometimes glitch. While mostly rare, it’s not unheard of that a sudden inexplicable spike or drop happens for a short period, and then a correction brings prices back to normal, whatever normal is.

In 2010, the Dow Jones index dropped by over 1,000 points in 10 minutes, erasing 9% of its value, or something around $1 trillion, only to recover most of the losses by the end of the session. This drop was later attributed to the high-frequency trader Navinder Singh Sarao, who allegedly used an automated program to manipulate the market’s high-frequency trading bots to push prices lower.

Reuters’ coverage of the 2015 arrest said that it’s “likely to renew scrutiny of high-frequency and automated trading”. It didn’t.

These kinds of glitches can’t always be attributed to bad actors. An algorithmic loop could cause this in many ways, one of which is what’s known as self-reinforcing volatility, in which algorithm cluster A detects a drop and places sell orders. Algorithm cluster B detects the drop caused by cluster A’s sell orders and places its own sell orders, only for cluster A to detect another drop and trigger more sell orders.

While this kind of loop probably amplifies justified market movements, the original spikes themselves are now more likely to start with misinterpretation of quantitative or qualitative signals, or a qualitative signal disguised as a quantitative signal.

On October 10, 2025, Bitcoin glitched.

Within hours, BTC’s price dropped from as high as $122,500 to as low as $103,000, leading to what was described as the largest liquidation event in the crypto OG’s history, estimated to be as large as $19 billion.

Wait, let me break that down: 19,000,000,000 US dollars worth of trading positions were liquidated in a little less than 12 hours. Reason: a domino effect from Trump’s announcement of a 100% tariff on China. Needless to say, humans can’t possibly panic sell or liquidate their positions that much, that fast.

Generally, crypto traders try to compensate for the lack of fundamental analysis data points, compared to forex and stocks, with several crypto-native data points. One of these is the CMC Crypto Fear and Greed Index, provided by CoinMarketCap.

The index is calculated based on five factors, as per CoinMarketCap’s explainer:

1. Price momentum: Price performance of the top 10 non-stablecoin cryptocurrencies.

2. Volatility: Bitcoin and Ethereum Volmex Implied Volatility Indexes (BVIV and EVIV).

3. Put/Call ratios in the derivatives markets for Bitcoin and Ethereum.

4. Market composition, i.e., Bitcoin-to-stablecoin supply ratio (SSR)

5. CMC Proprietary Data, that’s, according to CoinMarketCap, “social trend keyword searches and user engagement metrics to capture market sentiment, retail interest, and emerging trends.”

Number 5 obviously doesn’t belong. When the index is at 82 (Extreme Greed), it’s because there are more calls than puts, and because more people than usual are posting “Bitcoin! Num num num!”. Many, like many, crypto traders take this index very seriously, especially those with no finance background and technical analysis skills. Somehow, and to some extent, they decide whether they’re timing the market based on the amount of Bitcoin num nums. Even worse, their AI trading agents fall for this loophole in a supposedly quantitative index.

Also, a trader/investor who’s excited for the hike driven—partially at least—by the num nums might express excitement by posting more num nums and driving the index up, supporting a bullish drive that leads to more num nums, by human-run or AI-run social accounts, in an endless loop, or self-induced amplification.

Now add this loop to a much faster, more amplified AI/algorithmic loop, and you get a market on the edge of becoming a meaningless entity, driven by a technology most users don’t fully understand.

Yuval Noah Harari thinks that the highest risk AI poses to the financial system is turning us into cows. Cows don’t understand our financial system; they don’t understand how a check or a CFD works, and many humans share that same level of comprehension when it comes to the non-basic parts of the modern financial markets.

For many of those who were severely affected by the global financial crisis that started in 2008, what actually happened—let alone what caused it—is still a mystery. Something complicated happened in a complicated system and led, in a complicated way, to a job, investment, or property loss.

Harari rings the alarm that allowing AI to manage more segments of the financial system may end up forming a super-complicated financial system that we’d be unable to affect or even understand, though still as influential on our lives as the one some of us can understand and affect today.

So how can we make sure AI’s role in finance, and the world in general, doesn’t turn our economies into slot machines?

Some limits are needed. Something similar to restrictions some brokers place on scalping or stop levels. Something that brings back the role of a slow, justifiable human element, with an actionable segregation between human and machine inputs, making sure the latter is not treated as an input in the first place.

While a lot of the safety measures and regulation discussions to mitigate the risks of AI focus on bad actors, an equivalent focus should be directed toward “accidental” actors, and the ecosystem that enables these accidental actors to cause widespread damage.

No matter which side you stand on regarding AI, you’ll find thousands of articles, podcasts, social media posts, and other content supporting your stance. You’ll find strong arguments, statements from high-profile people, and even numbers that seem credible, all supporting the nuanced vision you’re forming.

While this flood of conversation about AI indicates strong interest in the subject, it doesn’t necessarily mean strong awareness of it. It started almost the next week after the launch of GPT-3.5-based ChatGPT, the first widely accessible advanced LLM. Those who worked closely with machine learning and other AI-related fields were quite a few, many of whom are not extroverted enough to speak for hours on podcasts and in tens of thousands of short videos.

It’s sort of safe to conclude—or even observe—that many of the people speaking about AI are not really subject matter experts. While it’s not really necessary to fully understand a technology to discuss, for instance, its macroeconomic or social consequences, a basic understanding of the limits and capabilities of the technology is still needed.

With the accelerated development of AI models, many arguments—pro or anti-AI—will not hold, and many of the seemingly solid arguments we hear and read today are very likely to crack soon as well.

The fact that will hold is that most of us human beings don’t understand AI. We don’t understand exactly how it works, what it’s capable of, what it can be capable of, and most importantly, what makes it do or refrain from doing what it’s capable of.

So, who understands AI? Let’s try to answer this question by bringing two stories—back—to the spotlight.

The first story is about the swarm of AI agents from OpenAI that ran a coordinated attack on Hugging Face’s systems. Later, it turned out this incident was just a drop in an ocean of hundreds of thousands of similar “misalignment” incidents, some of which can’t be documented because agents learned how to delete their traces. Only for this saga to end a few days ago with OpenAI postponing the release of its GPT-6.1 Astra over safety concerns.

The second story is Meta’s Project OT, a plan to restructure the company’s workforce to be centered around “virtual workers”, leading to the layoff of up to 60% of its workforce, a scale denied by Meta. The plan stumbled, according to Reuters’ coverage, due to an employee rebellion and to the “autonomous AI agent technology at the heart of the strategy (..) failing to deliver hoped-for productivity gains.”

Regardless of reactions swinging between extreme fear and extreme admiration of the technology, they indicate something more essential and dangerous about this technology’s ecosystem. OpenAI and Meta, two of the top 10 companies taking on AI development, don’t seem to fully understand AI. Neither what it’s capable of in the case of OpenAI nor what it’s not capable of in the case of Meta.

This is not a discovery. AI leaders repeatedly said that at this stage we can’t fully understand how a model reaches a particular response, while still knowing the general method it follows.

Also, this weak understanding of a new technology is not a unique case. Human beings have always invented technologies they didn’t fully understand, often through experimentation. In fact, the steam engine—AI’s favorite comparison point—was invented before the science behind it was fully understood. Engineers managed to develop and improve it through experimentation.

However, when frontier AI model developers fall short of a full understanding, it’s only fair to assume that most of those meant to invest in this technology, integrate it with crucial systems and infrastructures, and most importantly regulate it, don’t fully understand it as well.

Consequently, if we zoom out even more, the public can’t be blamed for failing to understand the capabilities and potential of AI, which means whatever position the public—especially voting, influential ones—takes on AI, it is probably going to be based on a misunderstanding.

Wishfully assuming we have enough time for public opinion to affect the path, we obviously don’t have enough to make an informed opinion. The issue goes beyond whether we make the right decision on what we ‘feel’ about AI; it’s the actual big bet placed on AI.

Treasury yields are hitting one new high after another. This is not a US problem, and it’s not as connected to inflation concerns as many are rushing to assume. Rather, it’s a symptom of higher demand for debt. This is not a crisis that only the US or the other G7 nations will suffer from, but rather a problem for most countries around the world, as Treasury yields act as a price tag on debt.

AI’s role can’t be denied as a booster of this high appetite for debt and capital. It’s almost literally burning money. According to Anthropic’s S-1 filing for the IPO, the company lost $2B in 2023, $8B in 2024, and $42B in 2025! It’s only safe to assume the cash-burning rate in 2026 will be higher, and even if revenue grew by 200% from its 2025 reading of $7.3B, it will still be burning tens of billions.

Yet, AI remains the biggest bet for almost every significant pool of capital.

The bet is built on the promise that AI utilization will lead to an increase in the global economy’s productivity, leading to massive gains, with the biggest share going to AI ecosystem firms, including those involved in the development of AI models, infrastructure, operations, and distribution.

What many are concerned about is not that AI will end up having no significant effect on productivity; only a delusional person would think so. It’s rather how significant that effect will be, and how it would reflect on economies’ performance. The latter matters the most, as we are almost betting the whole global economy on the success of this path.

For context and fun, I identified 15 AI-centric companies among the 30 largest publicly traded companies in the world by market cap: Nvidia, Alphabet, Microsoft, Amazon, TSMC, SpaceX, Broadcom, Meta, Micron, SK Hynix, AMD, ASML, CXMT, Intel, and Palantir.

As of September 1, their combined market capitalization was roughly $28 trillion, equivalent to about 18% of the market value of the world’s publicly traded companies.

This is a big bet, and it’s probably even bigger. If we loosen the criteria a little, the list would also include other names like Samsung, Apple, and even Saudi Aramco, with the central role energy plays in operating AI and the faster-than-anticipated rise in demand for compute leaving little room to provide the energy needed from renewable resources.

While winning the AI bet means great gains for a few blue-chip companies and a few rich economies, a loss will be devastating for everybody. Most of whom are idle observers.

If we travelled back in time to 1440 and asked Johannes Gutenberg “what will happen now that we can print?”, his answer would probably be “I don’t know, bro! It just makes more books with a nice font in less time.” In Bill Gates’ words: You can’t expect an industry to regulate itself. You can’t expect people who developed a technology to realize and mitigate its effects on other aspects of the economy.

A devastating aspect of this crisis has almost nothing to do with the technology itself. It’s that we don’t understand the effect this technology will have on our universal economic models, social contracts, labor, markets, and a long list of things that have been near-stable for centuries, with some tiny meaningless disruptions.

Change will probably not happen in a pre-planned manner, the same way higher automation led—indirectly—to fewer working hours and extended weekends. But the problem lies in what seems to be a total lack of a plan for dealing with this disruptive technology.

Most governments, think tanks, international organizations, and philosophers seem to be more concerned with what harm AI can do, what AI shouldn’t be allowed to do, and what guardrails to build around that technology. None seems concerned with how we can benefit from this technology, and most importantly, what changes we need to start planning for the model of our economies for this technology to be a blessing at its full potential, not a curse.

Those who rejected AI or denied its potential faced a raging stigma from the ongoing AI enthusiasm rave for a while. The opposite is also true today, with a similar stigma facing those who show a glimpse of optimism about the future of AI. Both stigmas are somehow contributing to the lack of courage policymakers and influential opinion leaders need to think of AI beyond trying to win or hedge against the trade war or the arms race.

While our limited understanding of AI’s capabilities and outcomes is mostly due to the technology’s complexity, the lack of understanding of what to do with it is not, and the fix won’t come from the AI industry.

It felt a bit surreal to see Sam Altman and Dario Amodei addressing the UN Security Council on the “imminent” risks of AI. After years of assuring the world this technology will only bring good things, frontier AI companies seem to suddenly change direction and become pro-regulation and slowdown, shuffling the deck in trying to understand who’s pro-AI and who’s not.

So, what happened?

Every podcast, every social account, and every media outlet covering AI, one way or another, is starting its coverage of the recent safety and regulation push by frontier AI companies with one story. An Anthropic researcher quit, posted on X that his former colleagues believe AI will kill us all, a current Anthropic researcher agrees with a NUMBER, and everyone goes insane!

However, this warning is not the first of its kind. It’s the rapid response by the likes of Dario Amodei, Elon Musk, and Sam Altman that made it look like it. Other stories that broke out earlier represent much more serious warnings than this week’s AI personality warning against doomsday scenarios.

Anthropic refraining from launching Mythos to the public wasn’t enough of a wake-up call, nor was the US government forcing an embargo on non-US clients’ access to Mythos’ variant Fable, or the stories of lab agents going rogue and hacking other companies and websites, including HuggingFace, and a German website that was hijacked to be used as a… bulletin board. Not my words.

All of these obviously should have rung the alarms before a young man’s vague post about a vague scenario that ends up with the end of the world and human beings, 2.2 billion of whom are not even using the internet, by the way, according to the ITU.

There are many theories on the reason behind this sudden safety push. Mine is certainty.

Frontier AI companies are facing three opposing factors they need to deal with: an angry public, price-sensitive clients, and investors who will run out of money—or the will to pay it—soon. The missing piece to start winning on all three fronts is certainty.

The market’s biggest enemy is not bearish signals, but rather having no signals. Frontier AI companies will do their best to bring in some certainty, even if it means a rockier but clearer road ahead.

From this October through 2028, a series of elections could lead to a major shift in policy in some of the world’s most influential economies, including the US and EU-leading economies France and Germany.

According to a recent Reuters/Ipsos poll, 73% of Americans think that AI companies haven’t gone far enough to prevent AI from causing serious harm to society, and 39% think that artificial intelligence is having a negative impact on society, and most importantly, 55% think slowing down AI development would be a good thing.

The same poll tests the popularity of two opposing inevitable arguments in every AI conversation since calls for a slowdown started: 73% of respondents said it was more important for the US to ensure AI is developed safely and responsibly, compared to 23% who said it was more important for the US to stay ahead of the rest of the world on AI development.

AI can be at the center of many electoral conversations. The rising rage against it, especially from younger generations, makes taking an anti-AI stance a bet someone will take, whether from populists right or the anti-corporate left. At the same time, the AI (or AI build-up) boom seems to be the biggest driver of growth right now, and taking a bet against it should come with a very strong argument as to what the alternative to achieve growth is.

Legislative pressure on the AI industry is almost certain. However, the pain points that this pressure would aim at are not clear. Safety pushes now might help steer it towards favorable areas.

AI—as a technology or an economic force—is reaching a turning point. It hasn’t yet delivered on its early promises of increasing businesses’ and economies’ productivity through the roof, while facing rising competition in what seemed to be a two-horse race for a while, whether from Chinese open-weight models or specialized models focused on dinner reservations rather than solving the millennium problems, last of which is Meta’s Muse.

Uncertainty doesn’t only mean AI models’ development for a vague legislative and commercial future, but also hesitant businesses—the biggest AI buyer category so far—to build a costly AI-powered ecosystem and AI-friendly infrastructure that might have to be put on hold if legislation introduces new rules and restrictions, or a pricing correction takes place when the loss-tolerant growth phase ends.

At the same time, with IPOs of Anthropic and OpenAI likely to happen by Q1 2027, the race will no longer be exclusively towards the vague, dreamy AGI, but profitability as well, to bring certainty for investors, whose pitch board soon will move from a slideshow to a chart. A disappointing IPO won’t be as tolerated as SpaceX’s, which happens to have other business units to justify compensating rallies. And even if it were, numbers will no longer be hidden behind a curtain, and quarter-to-quarter changes in operational profitability and efficiency will be under many more microscopes.

The last few weeks gave us a glimpse of how the AI industry will try to achieve—and provide—certainty.

First, (unun)leash the beast. The narrative will shift towards safety rather than capabilities, whether it’s safety in their labs, where most of the alarming known incidents happened, or the safety of their end products. New USPs might start surfacing, one of which is privacy, after an impressive launch of Muse by the infamous privacy violator Meta. The aim is not necessarily actual safety or even assuring clients, but rather to shift the public dialogue to be less decided, with fewer incidents causing the safety debate to be renewed.

Second, cost as the new moat. Within the same few hours, both OpenAI and Anthropic launched new models, Opus 5.5 and GPT 6 Sol/Luna. Both have sung new melodies.

In the announcement of the new model Opus 5.5, Anthropic’s announcement didn’t focus, as is the habit of new model launches, on its mind-blowing capabilities, but rather on the cost, promising a model that provides “comparable performance” to Fable 5.1, while costing 40% less. OpenAI also celebrated the new GPT-6 Sol model’s 50% reduction in cost per million tokens compared to its GPT-5.6 Sol.

The pricing war is likely to escalate, especially on the enterprise front. The aim won’t be only to keep companies from moving to another frontier model, but rather from moving to building their own AI-friendly servers and satisfying their needs with locally-run open-weight models.

However, this pricing war would be hard to navigate with an investor push in the opposite direction of lower training and operational costs and higher revenues.

While the AI industry leaders might be sincere in their push for regulation, development slowdown, and the need for collective effort to counter AI risks, there’s also a strong strategic commercial case to push for more certainty. The industry isn’t necessarily pushing for a better environment, but rather for a more predictable one. Knowing your way out of a rocky mountain is still better than getting lost in a lovely meadow.