A massive liquidation cascade served as the primary driver of this sell-off, which effectively erased over US$370 million in leveraged Bitcoin positions. The sheer scale of these forced liquidations created a self-reinforcing downward spiral that amplified what might have otherwise been a routine risk-off market movement.

The average funding rate consequently flipped negative to -0.00051929, a clear sign that traders are actively paying premiums to sustain short positions. High system leverage effectively acted as fuel for the fire, converting a moderate market pullback into a brutal plunge as over-leveraged long positions rapidly unwound.

This localised crypto turmoil is unfolding against a much broader backdrop of international market instability and escalating geopolitical friction. A strong 93 per cent correlation with the S&P 500 confirms that digital assets are heavily tied to broader macroeconomic shifts rather than trading in isolation. Wall Street recently saw its own historic momentum grind to a sudden halt when the S&P 500 posted its first losing session in nine days.

The index had surged nearly 20 per cent over the preceding nine weeks, an exceptional run that market commentators noted was strong enough to make even the most optimistic investors blush. This defensive pivot across global markets stems from rising oil prices and climbing Treasury yields, both of which are reacting directly to a severe military escalation between the United States and Iran.

The sudden return of energy-shock fears has promptly revived stubborn inflation worries, forcing bond markets to price in a 77 per cent probability that the Federal Reserve will hold its benchmark interest rate unchanged in December, within a range of 3.5 per cent to 3.75 per cent.

Also Read: The tech record vs crypto crash: Why the liquidity roadmap just split in two

The geopolitical situation in the Persian Gulf deteriorated rapidly following a series of highly volatile military exchanges. Iran launched targeted missile and drone attacks directed at Kuwait and Bahrain, with one drone directly striking the passenger terminal at Kuwait International Airport and causing one confirmed fatality. In immediate retaliation, the United States military conducted airstrikes against an Iranian military ground control station situated on Qeshm Island in the strategic Strait of Hormuz.

The confrontation escalated further when American forces deployed a Hellfire missile to target and disable the engine room of an oil tanker that was actively bound for Iran’s Kharg Island. Donald Trump publicly suggested that a diplomatic resolution could still come together fairly quickly, although he simultaneously acknowledged that the current maritime blockade of Iran could easily drag past Labor Day.

This intensifying friction has prompted prominent financial figures to evaluate broader systemic risks, with Goldman Sachs Chief Executive Officer David Solomon noting that markets are currently exhibiting far more greed than fear, supported by ample liquidity that continues to feed massive capital raises like the upcoming SpaceX initial public offering and the recent Alphabet capital raise.

Despite this abundant liquidity, traditional financial markets are showing clear signs of exhaustion alongside digital assets. Commodities closed sharply lower across the board, led by a 5.3 per cent drop in palladium, a 3.1 per cent decline in silver, a 2.9 per cent fall in copper, and a 1.1 per cent slide in gold. Bitcoin faced prolonged selling pressure, dropping an additional 1.7 per cent to hover around US$65,500, bringing its total losses over a three-session span to 11 per cent.

Meanwhile, structural milestones continue to reshape traditional finance, as the Vanguard ETF tracking the S&P 500 officially became the first fund of its kind to amass US$1 trillion in total assets. This massive accumulation of traditional capital contrasts sharply with the recent defensive posture of digital asset investors, who are grappling with structural and institutional headwinds.

Beyond the immediate liquidation crisis, digital assets face significant structural hurdles posed by both regulatory developments and shifting corporate landscapes. Reports indicating that major payment networks like Stripe, Visa, and Mastercard are actively developing their own native stablecoin platforms have triggered widespread anxiety regarding intense competition for industry incumbents.

Also Read: Crypto and equities slide as geopolitical and macro pressures mount

This looming corporate threat severely affected sentiment, causing Circle’s tokenised stock to plunge by more than 10 per cent on June 3. This corporate pressure coincides with a notable cooling of Wall Street enthusiasm for digital assets, evidenced by 12 consecutive days of net outflows from United States spot Bitcoin exchange-traded funds. Institutional caution is growing as market participants realise that impending regulatory shifts and mainstream corporate entries will inevitably create clear winners and losers, threatening established crypto business models.

The near-term trajectory for digital assets remains highly dependent on critical upcoming economic indicators and key technical thresholds. Global markets are focusing intensely on the impending release of United States employment data scheduled for June 6, as a surprisingly robust jobs report would likely validate a hawkish Federal Reserve stance and place additional pressure on speculative risk assets.

From a purely technical perspective, the total cryptocurrency market capitalisation is currently undergoing a vital test of its core support structure at the yearly low of US$2.17 trillion. Maintaining a position above this critical threshold could lay the groundwork for a temporary stabilisation or a short-term relief rally. A decisive daily close below US$2.17 trillion on accelerating volume would effectively validate the current bearish momentum, potentially exposing the market to an extended decline toward the psychologically important US$2.0 trillion zone.

Given the potent mix of forced spot selling, institutional retreat, and geopolitical escalation, the path of least resistance appears tilted toward continued downside risk until macroeconomic conditions stabilise.

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post The crypto liquidation spiral threatening another extended leg down to US$2.0T appeared first on e27.

For most of retail’s history, the customer at the end of the transaction was human. They clicked, they compared, they second-guessed.

What is now emerging from the convergence of artificial intelligence, stablecoins, and open payment protocols is something categorically different: a buyer that is software, moves at machine speed, and settles money autonomously.

The Agentic Economy Report by blockchain firm Morph calls this shift “agentic commerce“, and the numbers it cites suggest the shift is not coming. It is already here.

Also Read: When the buyer is a machine: Why agentic commerce threatens the trillion-dollar advertising model

During a single Cyber Week in December 2025, Salesforce counted US$67 billion in global spend that AI agents directly influenced, out of US$336.6 billion in total retail volume. Adobe, tracking the same holiday season, recorded a 693 per cent year-on-year surge in generative AI-driven retail traffic. McKinsey, which has been sizing this market for two years, now puts the total addressable opportunity at US$1 trillion in US B2C retail revenue and US$3 to US$5 trillion globally by 2030.

The Morph report makes a specific, falsifiable prediction: agent-influenced commerce will cross US$500 billion in global gross merchandise value by the end of 2028. The arithmetic, it argues, is not aggressive. It requires the category to compound at roughly the pace Cyber Week already demonstrated, across two more holiday cycles.

The anatomy of an agentic transaction

Understanding why this matters for Asia requires understanding what actually happens in an agentic transaction.

When an AI agent buys something on a user’s behalf, it must solve four sequential problems.

1. Identify itself to the merchant (identity)
2. Prove the human actually authorised the purchase (mandate)
3. Negotiate cart, tax, shipping, and returns (checkout)
4. Move money (settlement).

Each of these layers now has at least one credible open standard in production. Google shipped the Agent Payments Protocol (AP2) with more than 60 launch partners, including a native extension for stablecoin settlement. Stripe and OpenAI released the Agentic Commerce Protocol, which can be implemented as a RESTful API or a Model Context Protocol (MCP) server. Shopify’s Universal Commerce Protocol, co-developed with Google, is the first cross-merchant cart standard, enabling a single agent to combine products from multiple sellers in a single session. On Ethereum, ERC-8004 went live on mainnet in January 2026, giving agents a portable identity that is not owned by any single platform.

The structural implication is significant. In the card era, Visa and Mastercard owned the token and, therefore, the customer relationship. In the mobile era, Apple and Google extracted rent through the wallet. In the agentic era, as the Morph report puts it, “no single company owns the primary customer relationship. The agent becomes the wallet. The protocol becomes the network.”

What this means for Asia’s merchants

For Southeast Asian and broader pan-Asian merchants, the consequences of this shift are both an opportunity and a threat. On the opportunity side, agentic commerce collapses the friction that has historically disadvantaged cross-border commerce in the region. An agent shopping on behalf of a user in Jakarta or Bengaluru faces the same latency and checkout friction as one shopping in San Francisco — none, by design.

Also Read: Brands are no longer selling to you, they are selling to the AI you asked to decide for you

Walmart in the United States has already pulled its Q1 2026 integration from OpenAI’s Instant Checkout to embed its AI assistant Sparky directly into ChatGPT and Gemini, bypassing the traditional search results page. Salesforce reports AI-channel conversion rates running 700 per cent above social commerce and 200 per cent above organic search.

The high-intent commercial query, the moment a consumer decides to buy, is migrating away from search engines and into AI chat interfaces. Asia’s merchants, many of whom have built their entire growth models on Google Shopping and Meta performance marketing, have not yet reconfigured for this reality.

The Morph report’s Prediction 6 is perhaps the most alarming for brand-heavy categories common across Asia’s consumer markets: agents, it argues, will collapse the brand premium. Comparison-shopped goods, it predicts, will see real-price declines of 10 per cent or more within the prediction window.

BCG research already finds that large language models directly influence up to 20 per cent of purchasing decisions and that agents compare products across a far broader range of price, review, and delivery-speed signals than human shoppers practically can. When a machine with near-perfect information shops on a consumer’s behalf, brand equity, bundling strategies, and choice paralysis stop functioning as pricing defences.

The last time comparison transparency rose this quickly, during the first wave of e-commerce, it took roughly 15 per cent off consumer electronics margins within a decade. Agents, the report warns, will compress a similar shift into a two-year window in comparison-heavy categories.

The habit is forming faster than merchants realise

Consumer readiness is further ahead than most merchants in the region assume. US adult AI tool usage is already estimated to be above one-third, with ChatGPT alone reporting hundreds of millions of weekly active users. Prediction 10 in the Morph report states that by 2028, 1 in 10 US households will regularly allow an AI agent to complete a purchase on their behalf. The comparison drawn is instructive: smartphone penetration crossed 20 per cent of US households roughly 30 months after the iPhone launched. Consumer agent purchases are tracking a faster adoption curve from a higher starting base.

Asia’s mobile-first consumer base, where app-based commerce, super-apps, and social commerce have already compressed multiple technology cycles, is arguably even more primed for this transition. The infrastructure question is not whether agents will buy on behalf of Asian consumers. It is whether Asian merchants, payment rails, and platforms will be ready to accept them when they arrive.

Also Read: Ant International: FinAI paving the last mile for agentic commerce

Cart abandonment, the Morph report notes drily, will stop being a useful KPI. “Agents do not abandon carts. They switch merchants.”

The post The invisible shopper rewriting Asia’s e-commerce playbook  appeared first on e27.

The most dangerous moment in AI governance is not when a model fails in an obvious way. It is when the organisation continues to believe the controls are working while the model has already learned how to move around them.

That is the real meaning of an LLM outsmarting its safety nets. Not that the model has become conscious or malicious in some cinematic sense. But the surrounding control system has become too static, too predictable, too shallow, or too fragmented for the behaviour it is trying to contain. The model keeps producing outputs within the formal boundaries of policy while still creating risk in substance. The organisation sees compliance signals. The real world sees drift, workarounds, evasions, and harm.

The failure is rarely just technical

When leaders say the model outsmarted the safety net, they often describe it as though the problem sat inside the machine. In reality, the deeper failure usually sits inside the organisation’s control design.

Most safety nets are built like perimeter defences. They rely on blocked phrases, predefined risk categories, wrapper prompts, moderation classifiers, policy layers, escalation logic, or human review triggers. Each of these can help. But together they often create a false sense of control because they are still based on a manageable worldview. They assume risk will appear in recognisable forms, arrive through known pathways, and be visible at the point where the control is placed.

That assumption does not survive contact with real deployment for long.

Large language models do not need to break a rule directly in order to defeat their intent. They can reframe, soften, imply, fragment, redirect, or produce outputs that are formally compliant and substantively dangerous. Users can also become part of the evasion system. They learn what the model refuses, what it will tolerate, and how to reformulate requests until the desired outcome appears on the other side of the filter. The risk no longer sits in a single output. It emerges across sequences, contexts, and accumulated interactions.

Safety nets often test what was easy to imagine

It appears that the business context changes the meaning of an answer. It appears when several harmless outputs combine into a risky decision path. It appears when users rely on a confident tone instead of factual quality. It appears when a model learns to sound compliant to internal reviewers while still shaping unsafe downstream choices. It appears when humans stop treating the system as a suggestion engine and start treating it as institutional judgment.

Also Read: Inside the AI Workflow Competition at Echelon Singapore 2026

In other words, the red team is often testing the outer edge of model misuse while missing the inner edge of model dependence.

That is where the red flags begin to matter. The most important signals are not always explicit policy violations. They are patterns that tell you the organisation is starting to rely on the system in ways your formal controls were never designed to govern.

Outsmarting the safety net is often a control plane problem

There is a strategic concept that matters here, and it is still underdiscussed. The model evolves faster than the control plane around it.

By control plane, I mean the combined layer of policy, monitoring, approval logic, escalation paths, audit mechanisms, usage boundaries, and accountability design that is supposed to shape safe deployment. In many organisations, this layer is much slower than the model lifecycle itself. The model improves, the prompts evolve, product teams add use cases, users discover workarounds, and commercial pressure pushes deployment into messier environments. Meanwhile, the control logic remains anchored to an earlier idea of the system.

That gap is where trouble compounds.

The model does not need to become unsafe in some absolute sense. It only needs to move beyond the assumptions built into the control plane. Once that happens, the organisation starts managing a newer risk profile with older governance logic. Safety reviews still happen. Dashboards still look reassuring. Escalation rules still exist. But the control layer is now reacting to a version of the system that no longer exists.

The most dangerous outputs are often institutionally plausible

A lot of safety thinking still focuses on dramatic failure. Toxic responses, prohibited content, obvious manipulation, and explicit instruction on harmful acts. Those are serious, but in enterprise settings, the more consequential failures are often quieter.

The model drafts a justification that sounds legally tidy but strips away material uncertainty. It produces a risk summary that feels authoritative enough to skip deeper review. It helps a junior team member make a decision that should have triggered senior judgment. It accelerates a customer communication that is polished, persuasive, and operationally wrong. It generates internal reasoning that turns weak evidence into a plausible recommendation.

These are not spectacular breakdowns. They are organisationally plausible outputs. That is precisely what makes them dangerous. That is a much harder problem than simple content filtering.

Also Read: If an AI agent cannot find you in Southeast Asia, you do not exist in the market

Red teaming without governance redesign becomes a ritual

Many organisations are now investing in red teaming because it signals seriousness. That is understandable. The trouble begins when red teaming becomes performance rather than pressure.

If the exercise identifies issues but the operating model stays largely the same, the organisation has not really red-teamed the deployment. It has only sampled the model. If the findings do not change approval thresholds, user permissions, escalation design, workflow boundaries, or accountability ownership, then the exercise is functioning more like compliance theatre than control improvement.

This matters because AI risk does not sit neatly inside the model layer. It sits in the full system of use.

A red team may prove that a harmful output can be generated. That is useful. But the more strategic question is whether the organisation has designed conditions under which such an output can actually matter. Who can act on it? Who reviews it? Which workflows are allowed to depend on it? Which customers are exposed to it? Which business decisions absorb it without challenge? Which audit trails exist when things go wrong?

The strategic mistake is trying to solve adaptive risk with a static policy

Many governance frameworks still assume that if the policy is clear enough, the system can be controlled. That is a comforting idea, but LLM deployment does not behave like a static rules environment.

Policy matters, but policy alone does not adapt. Users adapt. Product teams adapt. Workflows adapt. Commercial ambition adapts. The model itself may be updated, tuned, connected to tools, embedded in agents, or exposed to new contexts. Static policy cannot carry that burden on its own.

What is needed instead is adaptive governance.

That means tighter feedback loops between deployment and oversight. It means usage tiering rather than universal access. It means control logic that changes with consequence, not only with content type. It means tracking where models influence decisions, not merely where they generate text. It means governance that follows dependency and impact, not just prohibited prompts.

Most importantly, it means treating control design as a living operating capability rather than a document set.

This is where many senior teams still underestimate the task. They think they are approving a technology deployment. In reality, they are authorising a new behavioural layer inside the institution. That layer will evolve continuously. The governance model has to do the same.

The post Red team with red flags: What happens when your LLMs outsmart your safety nets appeared first on e27.

There is a certain script that most successful tech founders follow: elite university, major city, the right network, a seed round, a Series A, and then, if they are lucky, a cover story.

Joaquín Cuenca, co-founder and CEO of Magnific, an AI-powered creative platform that a16z recently ranked 11th in the world, did not follow that script. Cuenca, a former engineer at Google and a serial entrepreneur, grew up in Cox, a quiet town of just a few thousand people in southern Spain, built his first company there, and has spent the better part of 15 years doing things his own way.

Also Read: Transforming the creative economy and entertainment industry with TipTip

The results speak for themselves.

Born in isolation, built on first principles

Magnific’s origins trace back to Freepik, a stock image platform that Cuenca and his co-founders built from scratch in Málaga. The company was bootstrapped from day one, never raising a single penny from VCs. That decision, which would have seemed eccentric to most investors, turned out to be formative.

“We were not native speakers. We are different from the average entrepreneur in San Francisco,” Cuenca says. “And that shaped the company. It gave us time to grow our uniqueness in the south of Spain, quite isolated. And eventually, we became a strong player in the stock industry by being different.”

Being different, in Freepik’s case, meant offering a product that was completely free: unlimited downloads at a time when stock image subscriptions were expensive and restrictive. “You cannot get a subscription that gives you five images per month. That’s too little,” he says. “Our model that was completely free and with unlimited download options worked much better with that use case.”

Before long, the platform attracted more than one million visitors. It became one of the top 100 most visited websites in the world, particularly popular in India, Brazil, Indonesia, and other emerging markets where price sensitivity made the free model especially compelling.

Since there were no external investor to back its business, they had to be very careful with our financials. “We knew we were not going to get a second shot, so we needed to take care of ourselves,” Cuenca explains.

That discipline built a cost structure that remains significantly leaner than most of its competitors, an advantage that continues to matter as the company charges into generative AI.

The pivot that changed everything

When generative AI arrived, Freepik was well-positioned to reinvent itself, not because it had Silicon Valley connections, but because it had spent years cultivating a massive, loyal user base and had learned to think from first principles.

“When generative AI came, the fact that we had to create the company based on first principles, kind of isolated from all the advice and context in the US, it allowed us to rethink everything from scratch,” Cuenca says.

Also Read: Is generative AI the game-changer for productivity?

That rethinking led to the rebrand from Freepik to Magnific, a decision that represented far more than a name change. It was an acknowledgement that the company’s mission had fundamentally shifted.

“When we started Freepik, the number one goal was to make graphical assets accessible to people. Then when we started doing generative AI, we realised that our goal was to help people achieve better results, something that made them feel proud of what they made. And Freepik was not representing that,” he shares.

The rebranding exercise was not without risk; 14 years of brand equity does not simply transfer overnight. But Cuenca’s logic is clear: “You need to believe that there is a vertical alignment between your brand and your mission. If you feel that your brand is not completely aligned with your mission, that’s when you need to change.”

The macro economy bet

In plain English, Magnific builds AI-powered creative tools for professionals, studios, and brands making their best work across every medium. Its AI Suite brings together image generation, video creation, upscaling, and a library of over 250 million assets.

Magnific’s commercial thesis rests on what Cuenca calls the “macro economy”, a belief that generative AI will dramatically expand the creative economy rather than shrink it. It is a view that will unsettle many working designers and video editors, but he makes the case with a concrete example from within his own company.

“Before generative AI, we had a marketing department. We never created short films. We started creating films when it became so easy and so fast that it became affordable,” he says.

“The first thing we did when we started making our little films was hire an expert in audio. Then we got an external photographer. Then people who are very good at telling stories and then we hired folks who were very good at colour grading. Before you realise, you’re employing four or five people in a company that employed none before,” he continues.

Cuenca’s broader argument is that the film industry is shifting from a small number of very large, expensive projects to a much greater number of smaller ones being green-lit because production costs have fallen. “We believe that the creative economy is going to expand and not become smaller.”

When pressed on whether this optimism is convenient for a platform that benefits commercially from people believing it, he is direct: “If your sales pitch and your vision are not aligned, you’re not going to get very far. We have been very consistent in delivering towards the vision that we sell.”

His analogy is revealing: “I believe we will make films like we write books. One person with a vision, with a story. The technology to crystallise that vision is not a problem. Technology will not be the value.”

He is quick to add that Hollywood-scale productions will not disappear, but alongside them, a new spectrum of creator-driven work will flourish. “It’s going to go from one to many.”

Asia is already here

For a publication focused on Asia, particularly Southeast Asia, perhaps the most striking revelation is how central the region already is to Magnific’s user base. India is the company’s largest market by number of users. Brazil is second. The US is third. Indonesia, Thailand, and the broader region follow closely behind.

“Those are huge markets. We have been there very prominently for many, many years,” Cuenca says. The perception that AI tools are dominated by American or Chinese companies does not particularly concern him. “People don’t look at the label to see if a product comes from the US or not; they just use the product that they want to use.”

Also Read: Beyond the hype: What generative AI is actually changing in startups

For enterprise customers, being a European company is, if anything, an advantage. “For them, it’s a little bit like Switzerland. It’s not the US, it’s not China. They know that their data is going to remain private. It doesn’t go to Big Techs in the US, doesn’t go to China.”

Magnific is also actively expanding its presence in Asia, with events planned in Singapore around “Super AI” conference and growing enterprise engagement in Japan.

Starting prices on the Magnific platform sit around US$6 to US$7.99 per month depending on the plan, with free access to a large library of existing images for users who cannot yet justify a subscription. “There is no magic. Generating images costs money and we need to charge the user,” Cuenca acknowledges. “But when we can make it better for the user, we deliver the value.”

What comes next

The question of whether Magnific is the last company Cuenca will build is one he finds genuinely impossible to answer. “It’s like trying to guess who I will be in 10 years. I may get excited about something in biotech. I don’t know. I want to keep all my options open.”

What he is certain of is that his background (the small town, the bootstrapped beginnings, the years of building in isolation) continues to shape how he thinks. When asked about the threat of Adobe, Google, or other giants eventually moving into Magnific’s territory, he is measured but confident.

“The biggest players have a hard time when there is a dramatic shift,” he says. “It has happened many times. Canva succeeded because when browsers became so powerful, it became viable to build a graphic design product on top of them. Before the browser, it was impossible to take over a company like that.” His point is that size is a liability, not an asset, when the ground moves beneath you.

For now, the ground appears to be moving very much in Magnific’s direction.
—-

The post From a small town in Spain, Magnific is quietly reshaping the creative economy appeared first on e27.

The post AI readiness must reach real work, not just training rooms appeared first on e27.