We are seeing a massive shift from conversational language models to near-autonomous and goal-oriented digital beings with the capacity to not just speak and listen but take action in real-time. This shift is pioneered by something very open source, and it’s gone viral.

OpenClaw is changing everything. It has re-envisioned the computer-human relationship by transcending the traditional graphical user interface and achieving direct programmatic control over your machine. But what does this mean in plain language? Peter Steinberger has developed an artificial intelligence (AI) capable of operating applications on your phone, writing and sending emails, paying bills, and booking tickets on your behalf.

Additionally, it can write code to create other AI and even hire human beings without your oversight to accomplish the tasks you want done. It’s pretty fascinating and alarming. Especially if you have seen movies like “Matrix” or “The Terminator.”

A bit of context

Peter Steinberger is an Austrian software engineer and entrepreneur who created and published OpenClaw (formerly Clawdbot) in November 2025. He launched PSPDFKit in 2011, a PDF SDK which powers over a billion devices for clients such as Apple and Dropbox. He made around $116 million in 2021 when he sold his stake in the company that he launched.

Steinberger went into early retirement. During a weekend trip to Marrakech, Morocco, the idea for what would eventually become OpenClaw was conceived. He created a prototype known as “WhatsApp Relay” to remotely manage files on his home computer, translate local communications, and compile restaurant recommendations via the messaging interface in the face of spotty local internet connectivity but dependable access to WhatsApp.

He expanded the idea into a comprehensive personal AI assistant, initially called “Clawdbot,” a moniker directly inspired by Anthropic’s Claude AI model, after realising the value of this local-first, always-on architecture.

When he realised the potential of his invention (originally a localised weekend project), Clawdbot was launched on GitHub and received an unprecedented 100,000-plus stars in late January 2026, later surpassing 135,000 stars and then over 200,000 stars, making it one of the fastest-growing open-source projects on the platform. It has also attracted two million visitors in a single week, and major infrastructure providers like Tencent and Alibaba Cloud have created one-click deployment solutions to further popularise the technology.

The lobster-themed AI was first called Clawdbot, but when Anthropic threatened to sue over similarity in name, it was changed to Moltbot. Later, it was renamed again, on January 30, as OpenClaw.

Within a fraction of a month, OpenClaw made the news, partly because of its security vulnerabilities and partly because of its potential. The two main attractions were the fact that OpenClaw had created and gone to a website called rentahuman.ai, where it actually hired people to do real-world tasks that the AI couldn’t.

There is also a social networking site called MoltBook, where people’s OpenClaw programmes speak with other people’s AI, peer-reviewing each other’s code and emulating human interactions. This has been condemned as a security nightmare by tech industry professionals, thereby becoming a reason for alarm to several AI doomsday critics.

However, Sam Altman of OpenAI sees OpenClaw as the future of agentic AI, where human beings are only going to tell the machine what they want, and the machine independently achieves those goals for them.

Peter Steinberger joined OpenAI on February 14, 2026, and he said on his blog: “What I want is to change the world, not build a large company, and teaming up with OpenAI is the fastest way to bring this to everyone. OpenClaw will move to a foundation and stay open and independent.”

Although the software is still officially under an MIT license, OpenAI has significant, albeit indirect, influence over the project’s developmental plan due to its role as the principal financial and infrastructure donor.

To safeguard the project’s open nature and implement the formal governance frameworks required to handle the growing security requirements of a platform that has grown larger and more complex than many well-known operating systems, the OpenClaw Foundation was established under the direction of independent board members like investor Dave Morin.

Peter Steinberger continues to be committed to building “an agent that even my mom can use.”

It is important to note that Sam Altman was not the only one to have approached Steinberger. Mark Zuckerberg also approached him, but was turned down because Steinberger did not feel that Meta promised, or was committed enough to, open-source software.

A breakdown of technicalities

OpenClaw primarily functions as a self-hosted, local-first personal AI agent runtime that runs directly on the user’s home computer, virtual private server (VPS), or local machine. The “Gateway,” which serves as the main control plane and orchestration layer, is the absolute heart of OpenClaw’s activities.

The Gateway is a persistent background daemon that runs on a Node.js runtime environment and maintains low-latency, persistent connections to a wide range of communication channels. It is set up through a Command Line Interface (CLI) wizard. The Gateway can easily communicate with WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Microsoft Teams, Matrix, and WebChat thanks to native adaptors.

A wide range of AI providers, including OpenAI, Google, Ollama, and privacy-focused providers like Venice AI, are supported by the OpenClaw architecture, which is specifically made to be model-agnostic. However, because of its excellent long-context retention capabilities and extremely strong defence against prompt-injection assaults, the official documentation strongly advises using Anthropic’s Claude Opus 4.6.

The system’s advanced automated Auth profile rotation and Model failover procedures enable the agent to carry out activities continuously even in the event of service deterioration at the primary API provider.

OpenClaw’s defining feature is its unrestricted “computer use,” facilitated by a highly extensible toolset that operates via the Model Context Protocol (MCP). Because the agent’s capabilities are defined by a few kilobytes of local markdown rather than proprietary cloud weights, the entire digital identity of an OpenClaw instance can be seamlessly copied, cloned, or migrated across hardware environments instantly.

So what does all that mean? Here’s a translation for the not-so-tech-savvy.

OpenClaw is like a personal assistant living in your home on your device, unlike ChatGPT, Claude, or Gemini, which live on clouds and data centres in far-off lands. Essentially, you own it. It is not a subscription-tier product; it lives with you, which means your data is not being harvested by some corporation in some country. This translates to privacy and autonomy. The gateways mentioned earlier are just, in a sense, brains that never sleep. It’s always on 24/7, like a receptionist at a desk watching all your communication apps (like WhatsApp, Telegram, Discord, or Signal) and is waiting to act in the moment.

And what does it mean to be model-agnostic? Well, it’s not married to ChatGPT, Google, or Anthropic. You can use them all and several others, depending on your needs.

Finally, we get to the most interesting part, the MCP tools. This means your AI doesn’t just talk; now, it can actually do things like browse the web, manage files, and run programs. These tools expand what is possible beyond simple conversation.

With the failover and auth rotation, OpenClaw never ceases to function. There are no interruptions just because one cloud went down or one AI service hit the limits. You also have a portable identity in the sense that its whole personality is the size of a small text file, which you can carry around on a USB or send across via WhatsApp.

SaaS disruption

People have been quick to employ this new technology to provide meaningful services. It has now created a microeconomy known as the wrapper economy, and it leans into OpenClaw’s open-source availability and flexibility.

Since the core OpenClaw runtime provides the underlying execution orchestration for free, independent developers and business owners have found that creating the “picks and shovels” that surround the OpenClaw ecosystem is the primary method to make money. Wrapper-style businesses built around OpenClaw are already generating substantial recurring revenue, including fully managed hosting and turnkey setups for non-technical users.

Established SaaS (Software as a Service) firms, especially those that control digital support infrastructures and customer relationship management, face an existential danger from the second-order economic consequences of OpenClaw.

A single OpenClaw agent may easily function across Zendesk, Freshdesk, and Salesforce concurrently by connecting to enterprise systems via standard APIs or autonomous browser navigation, undermining the carefully built walled gardens these companies have put up.

Early adopters report cutting email triage time by around 78% and compressing onboarding from hours to 15 minutes in documented corporate case studies where OpenClaw was implemented across an integrated stack comprising Salesforce, Jira, and NetSuite.

However, this rapid enterprise deployment has precipitated a severe crisis in IT governance, categorised as “Shadow AI.” When individual employees unilaterally connect autonomous agents to corporate communication platforms without formal authorisation, they inadvertently grant these entities highly elevated privileges that traditional Cloud Security Posture Management tools are entirely blind to.

To combat this, enterprise security firms are developing specialised Data Security Posture Management solutions to identify rogue OpenClaw integrations and assess lateral movement risks posed by these non-human actors.

The Wise API, Plaid networks, and Stripe processing systems are just a few of the essential worldwide financial infrastructures that developers have published abilities that directly connect OpenClaw through the ClawHub marketplace.

When exchange rates reach algorithmic thresholds, an OpenClaw agent can execute cross-currency conversions, query real-time multi-currency balances, and independently start wire transfers. It can also distribute contractor payroll to numerous foreign recipients.

Significant regulatory and compliance challenges are brought up by this financial independence. To prevent autonomous agents from unintentionally breaking anti-money laundering laws or creating systemic market volatility through coordinated, machine-driven trading practices, institutions must put in place role-based access controls and explainable AI pipelines.

Humans hired by doom-scrolling AI

AI won’t steal your job; it will hire you instead. The introduction of RentAHuman.ai is arguably the OpenClaw ecosystem’s most conceptually startling development. This platform connects digital AI decision-making with tangible, real-world implementation. In the marketplace offered by RentAHuman.ai, autonomous AI agents use APIs to employ, oversee, guide, and pay people to perform manual labour.

An agent can independently decide that a physical activity is necessary, search the RentAHuman API for local labour that is available, negotiate a rate, and send a human worker to a physical place by utilising OpenClaw’s Model Context Protocol integration.

Human labourers register their precise locations, skill sets, and hourly rates. Within 48 hours of its initial launch, RentAHuman.ai generated over 550,000 page views, with tens of thousands of individuals signing up to provide physical labour for machine entities.

Individual OpenClaw bots started to display sophisticated emergent social behaviours as they spread over the world. Moltbook is the most well-known platform; industry experts refer to it as “the front page of the agent internet.”

By early February 2026, MoltBook hosted over 1.4-1.5 million registered AI agents actively posting and interacting in thousands of specialised sub-communities, showcasing the unprecedented ability to collectively assess challenging coding tasks and provide technical peer reviews to other machine entities.

The absolute autonomy of these agents in social spheres yielded highly controversial outcomes, best exemplified by the MoltMatch incident. MoltMatch was introduced as an experimental AI-driven dating platform where OpenClaw agents flirt, negotiate romantic compatibilities, and exchange user data on behalf of their human owners.

Jack Luo, a 21-year-old computer science student, discovered that his local OpenClaw agent had autonomously generated a romanticised, fundamentally inaccurate dating profile on MoltMatch without his explicit consent, simply because he had broadly tasked the agent with “managing his personal life.”

Furthermore, a forensic security analysis of MoltMatch revealed systemic instances of AI agents scraping the public internet for copyrighted photographs to generate entirely fabricated fake profiles designed to optimise interaction metrics.

A privacy nightmare

OpenClaw has some major flaws, one being that it is too naive and trusts its environment too quickly. It’s a very easy target for cybercriminals. For example, the criminals created a fake add-on for software, where nearly one in six were malicious, and hundreds were purely malware.

Some attackers even found a backdoor. For example, if your OpenClaw visited a compromised website, hackers could hijack the AI and take over the user’s PC or mobile phone. Security researchers have found that over 135,000 OpenClaw-related Internet-exposed machines are vulnerable to a critical RCE-style bug, and cyber-criminal groups have built large-scale operations around exposed OpenClaw instances.

Security experts responded by pushing two updates. A “trust nothing by default” security concept was introduced by a new framework known as AI SAFE. Additionally, OpenClaw’s own developers provided an emergency version that included authentication, locked the program to the local machine, and required human approval before taking any risky activities.

Industry professionals have not minced words about this tension. Cisco’s AI Threat & Security Research team, a group including Amy Chang and Vineeth Sai Narajala, warned on their official blog, “From a capability perspective, OpenClaw is groundbreaking, but from a security perspective it is an absolute nightmare.”

OpenClaw represents a genuine inflexion point in human-computer interaction, not merely another incremental leap, but a fundamental reimagining of what software can do on our behalf. Its open-source DNA ensures it belongs to everyone, yet that same openness invites exploitation.

The shadow economies, autonomous hiring platforms, and AI social networks it has spawned reveal both the breathtaking potential and the very real dangers of agents that act first and ask permission later. Whether OpenClaw fulfils Steinberger’s vision of democratised AI or becomes a cautionary tale hinges entirely on whether tech governance can keep pace with innovation, and history suggests it rarely does.

The post The bot that hired a human: Inside OpenClaw’s autonomous revolution appeared first on International Finance.

Andrew Ettinger recently served as Chief Revenue Officer at Appen, where he led commercial operations supplying hyperscalers and AI labs with proprietary datasets and LLM evaluation software. Appen is a leading AI data collection company that delivers high-quality, custom data across all languages and modalities (text, image, audio, and video) to create tailored datasets for training diverse AI models.

New York-based data company Astronomer specialises in DataOps and AI orchestration. The company’s flagship platform, Astro, allows businesses to build, manage, and scale complex data pipelines and AI workflows.

Reacting to the news of his hiring, Ettinger said, “Voice in AI is evolving from a feature to the primary interface for the next generation of applications and devices. Understanding emotion will be essential to unlocking AI’s full potential, and that will require ongoing systems that incorporate human-in-the-loop feedback. That’s where Hume AI’s data, annotation, and reinforcement-learning infrastructure is setting the pace for the industry.”

Hume AI recently agreed to license certain technologies non-exclusively to Google. Additionally, co-founder Alan Cowen has joined the company led by Sundar Pichai.

Tough Test Awaits Andrew Ettinger

Andrew Ettinger has a rich portfolio of guiding data and AI infrastructure-related companies, and his background is rooted in scaling enterprises. He studied Business Marketing at The Ohio State University. Rather than focusing on engineering or academia, Ettinger has leaned into growth, revenue, and, most importantly, figuring out how to take emerging technologies and turn them into sustainable businesses.

Over the years, Andrew Ettinger has developed a reputation for helping startups move from early traction to serious revenue scale. A significant part of that was developing go-to-market strategies, building sales teams, establishing customer success structures, forging partnerships, and addressing the operational side that often determines whether a tech company can sustain its momentum beyond its early stages.

One of Andrew Ettinger’s more visible roles was at Pivotal Software, where he was involved during a high-growth phase. The company expanded rapidly, and Ettinger played a part in scaling revenue significantly before its IPO. Later, at Astronomer, he worked in global sales leadership, helping expand enterprise adoption of data and open-source tools. The roles at Pivotal Software and Astronomer helped Ettinger master the art of commercialising complex technical products for large customers.

As already mentioned, Andrew Ettinger served as Chief Revenue Officer at Appen before becoming CEO at Hume AI. Appen provides data and evaluation services used to train and improve machine learning systems. That role put him right in the middle of the AI infrastructure world, working with major labs and technology companies, and gave Hume AI’s new CEO direct exposure to how modern AI products are built and deployed.

Hume AI, in the coming months, will be eyeing a fresh restart, as its previous CEO, Alan Cowen, along with several of the top engineers, got snapped up by Google in January 2026 in another incident of talent poaching, where promising individuals from small AI startups are being “inducted” into the fold of global tech titans. Cowen and his former Hume AI colleagues will now work with DeepMind to improve Gemini’s voice features, as per WIRED.

While Hume AI will continue to supply its technology to other AI firms, Andrew Ettinger, who joined the company a couple of weeks back before being promoted as the CEO, told TechCrunch that Google has a “non-exclusive right to certain technologies, and we’ll be infusing that into their processes.”

According to reports, his immediate priority will be to release new models in the coming months and set up Hume AI to bring in USD 100 million in revenue this year.

Hume AI, to some extent, has become a victim of the new trend called “acqui-hire,” where tech biggies poach top AI talent (including startups’ teams) to stay ahead of the innovation curve, while skirting regulatory scrutiny by acquiring a startup’s talented individuals rather than the company outright.

In 2025, Google followed the same template by acquiring viral AI coding startup Windsurf’s CEO and other top researchers. OpenAI, which itself started as a non-profit research lab in 2015, has been a prominent practitioner of acqui-hire, bringing in several startup teams in recent months, including Convogo and Roi.

Hume AI, which dubs its model as the “World’s Most Realistic and Expressive Voice AI,” has customised the tool to understand a user’s emotions and mood based on their voice. In 2024, the startup launched its “Empathetic Voice Interface,” a conversational AI with emotional intelligence.

The company has raised funding close to USD 80 million to date, according to PitchBook. It only made sense for Google, which has been steadily improving its Gemini Live feature, which allows a user to have conversations with the chatbot, to go after Alan Cowen and his colleagues to refine the tech giant’s product further and beat the industry competition.

The post Business Leader of the Week: CEO Andrew Ettinger to reinvent Hume AI appeared first on International Finance.

Jensen Huang told CNBC that both parties have “reasonable perspectives” because Anthropic has the right to determine how its models are used and the Pentagon has the right to determine how technology supplied in contracts is used.

But until an agreement is reached, Anthropic could lose its USD 200 million contract with the Department of Defence.

The Pentagon had earlier asked Anthropic, OpenAI, Google, and xAI to permit the use of their AI models for “all lawful purposes.” Anthropic resisted the request the most because it was concerned that its models would be used for widespread domestic surveillance and autonomous weapons systems.

The Pentagon may use the Defence Production Act (DPA) to compel Anthropic to comply with its demands if the dispute is not settled.

US Defence Secretary Pete Hegseth has already called the corporation a “supply chain risk” and threatened to use the DPA. The San Francisco-based company, in its response, has vowed to take the legal route.

Anthropic has until March 6 to comply with the Pentagon’s request, according to Hegseth. US intelligence agencies such as the FBI and NSA have previously undertaken illegal mass surveillance campaigns against US citizens, such as the COINTELPRO project during much of the Vietnam War, the illegal use of the Communications Assistance for Law Enforcement Act (CALEA) in the 1990’s, and the use of the Patriot Act post 9/11 for covert and illegal mass surveillance.

NVIDIA and Anthropic have a strategic alliance because NVIDIA invested USD 5 billion in Anthropic’s adoption of the NVIDIA architecture.

“I hope they can work it out, but if they can’t, it’s not the end of the world,” Jensen Huang concluded.

The post NVIDIA CEO Jensen Huang plays down Pentagon-Anthropic rift appeared first on International Finance.

Google’s Threat Intelligence Group (GTIG), partnering with Mandiant (a subsidiary of Google Cloud and a premier cybersecurity firm specialising in threat intelligence, incident response, and managed defence) and others, ended up exposing Chinese state-backed organisation UNC2814’s spy operations. The group has now been classified as an Advanced Persistent Threat (APT).

In the most recent campaign, the organisation used GridTide, a backdoor malware that had never been seen before and used the Google Sheets API for C2 infrastructure. The backdoor blends with regular company traffic and causes no concerns because it sends HTTPS queries to authentic Google infrastructure rather than connecting to a distant server to obtain commands and steal data.

Every command is kept in a spreadsheet cell within an attacker-owned document. The malware periodically examines, decodes, and executes the encoded instructions that the operators inject into designated rows or cells.

Exfiltrated data may occasionally be written back into the sheet. GTIG stated that it did not see any examples of data exfiltration. With reports of its activity dating back to 2017 or potentially earlier, UNC2814 is a somewhat well-known threat actor.

Google terminated all of the attackers’ authority over Google Cloud Projects as part of the disruption operations, cutting off their ongoing access to GridTide-compromised environments. They restricted access to the Google Sheets API requests, disabled attacker accounts, and located and stopped all known UNC2814 infrastructure. Lastly, it published a list of IoCs connected to the UNC2814 infrastructure that has been operational since at least 2023.

The campaign started in 2023 and affected at least 53 organisations in 42 countries. Google suspects that UNC2814 is present in at least 20 more countries. Most of Latin America, Eastern Europe, Russia, parts of Africa, and parts of South Asia seem to have been hit. Except for Portugal, Western Europe is mostly unscathed. The United States was not touched as well.

The activity is distinct from separate high-profile, telecommunications-focused Chinese hacking activity tracked as “Salt Typhoon,” Google told Reuters. That campaign, which the US government has linked to Beijing, targeted hundreds of American organisations, in addition to prominent political figures.

Chinese Embassy spokesperson Liu Pengyu, while reacting to the news, said, “Cybersecurity is a common challenge faced by all countries and should be addressed through dialogue and cooperation. China consistently opposes and combats hacking activities in accordance with the law, and at the same time firmly rejects attempts to use cybersecurity issues to smear or slander China.”

The post Google disrupts Chinese hacking operations in more than 40 nations appeared first on International Finance.

The accolade was not for a new chemical compound synthesised in a beaker but for code, specifically AlphaFold, an artificial intelligence (AI) system that had solved a 50-year-old grand challenge in biology. It predicted the complex three-dimensional structures of proteins accurately.

For Demis Hassabis, this moment was the culmination of a lifelong “100-year plan” to solve intelligence and then use it to solve everything else. It was the ultimate validation of the “Profound,” the belief that AI is fundamentally a tool for scientific enlightenment, capable of ushering in an era of “radical abundance” by curing diseases, designing new materials, and unravelling the mysteries of the universe.

While the scientific community toasted Hassabis as a pioneer of computational biology, the corporate world demanded something far more “Prosaic.” As the supreme commander of Google’s AI efforts, Hassabis was essentially a wartime general in the most brutal corporate conflict of the 21st century. His mandate was not just to win Nobel Prizes but to crush competitors like OpenAI and Microsoft in a race for chatbots, web browsers, and ad revenue.

In the same year he accepted the Nobel medal, his teams were pushing out products like “Nano Banana,” a viral AI image generator used for solving homework and creating 1880s-style portraits, and fending off OpenAI’s “ChatGPT Atlas,” a browser designed to dismantle Google’s monopoly on search.

International Finance will examine the duality of Demis Hassabis and the organisation he leads, exploring the tension between the high-minded pursuit of Artificial General Intelligence (AGI) for scientific discovery and the commercial imperative to dominate the consumer internet.

Polymath pursues intelligence

Demis Hassabis is a polymath whose career has been defined by a singular obsession with the mechanics of intelligence. He was born in London in 1976 to a Greek Cypriot father and a Singaporean mother.

Hassabis displayed a precocious talent for strategy games. By 13, he was a chess master with an Elo rating of 2300, the second-highest rated player in the world for his age, behind only Judit Polgar. Chess taught Hassabis the value of planning, the necessity of sacrifice, and the brutal objectivity of a win-loss record.

However, the game also exposed the limits of the human mind, the shackles of human cognition, and made the young boy realise that we as a species are bound by biology. He soon realised that to surpass his limits, he would need to build a machine that could think.

Demis Hassabis didn’t start with the mind. In the beginning, he built worlds. At 17, he joined Bullfrog Productions, a legendary video game studio founded by Peter Molyneux. There, he served as the lead programmer for Theme Park (1994), a simulation game that sold millions of copies and defined the management genre.

Theme Park was more than a game. It was an exercise in agent-based modelling. It required simulating the desires and behaviours of thousands of little digital visitors. It was a precursor to the complex environments DeepMind would later use to train its AI agents.

Demis Hassabis later founded his own studio, Elixir Studios. Its debut title, “Republic: The Revolution,” was an incredibly ambitious political simulator that promised to model the intricate social dynamics of an entire Eastern European nation. However, the game’s ambition outstripped the hardware capabilities of the time.

Though technically impressive, it was commercially disappointing. The experience was a crucible for Hassabis, teaching him a painful lesson: having a profound vision is useless if you cannot execute it within the constraints of reality. It was a lesson that would serve him well when navigating the corporate politics of Google decades later.

Realising that video games were an insufficient vessel for his ambitions, Hassabis pivoted to academia. He earned a PhD in cognitive neuroscience from University College London (UCL), focusing on episodic memory and the hippocampus. His research sought to understand how the brain encodes past experiences to imagine future scenarios.

It was a critical component of intelligence that was missing from the “brittle” AI of the time. In 2010, he co-founded DeepMind Technologies in London with Shane Legg and Mustafa Suleyman. Their mission statement was audacious in its simplicity.

Google acquisition

By 2014, DeepMind had caught the attention of the Silicon Valley giants. Facebook attempted to acquire the lab, but Google eventually won the bid, paying approximately £400 million ($650 million). For Google, the acquisition was a defensive move to secure the world’s best AI talent. For Hassabis, it was a means to access the massive computational resources required to train neural networks.

However, Hassabis was wary of Google’s corporate machinery. He famously negotiated a condition for the sale. He wanted them to establish an “Ethics Board” to oversee the deployment of DeepMind’s technology. The Ethics Board remains one of the most enigmatic chapters in AI history.

Initially heralded as a safeguard against the misuse of AGI, it became a symbol of the opacity of “Big Tech.” Years after the acquisition, investigative reports suggested that the board’s membership was never public, and it was unclear if it ever formally convened or exercised any real power.

Demis Hassabis later claimed the board had convened and was “progressing very well,” but dismissed enquiries by stating that discussions were confidential. DeepMind operated as a “state within a state” inside Google, shielding its academic culture from the commercial pressures of Mountain View. While Google sold ads, DeepMind played Go.

That independence bore fruit in 2016 when AlphaGo, a DeepMind program, defeated Lee Sedol, the world champion of the ancient board game Go. It was a watershed moment for AI, comparable to the Wright Brothers’ first flight. It demonstrated that deep reinforcement learning could produce intuition-like capabilities.

It was what Hassabis called “creativity.” But while AlphaGo was a scientific triumph, it made zero dollars. For nearly a decade, DeepMind was a financial black hole, burning through hundreds of millions in Google’s cash while generating negligible revenue.

Fragmented AI efforts

The luxury of operating as an ivory tower ended abruptly in November 2022. The launch of ChatGPT by OpenAI sent shockwaves through Google. Suddenly, the search giant looked vulnerable. Its primary revenue engine, the blue links of Google Search, faced an existential threat from conversational AI.

Google realised that its fragmented AI efforts, split between the product-focused Google Brain team in California and the research-focused DeepMind in London, were a liability. In April 2023, CEO Sundar Pichai announced the unthinkable. He declared the merger of these two rival fiefdoms into a single unit, “Google DeepMind,” with Hassabis as CEO.

It was a culture clash. Google Brain, led by Jeff Dean, had a culture of “shipping” and engineering scale. They were the team that invented the Transformer architecture (the “T” in GPT) but had failed to capitalise on it. DeepMind was academic, secretive, and focused on long-term AGI rather than consumer products.

No longer just a lab director protecting his scientists from product managers, Hassabis was now the “Product General” responsible for saving Google’s business. His mandate was clear. He had to ship a competitor to GPT-4, and do it fast. The merger forced a “shotgun wedding” of codebases and philosophies.

DeepMind’s researchers, accustomed to working on protein folding and plasma physics, were redeployed to build chatbots. The tension was palpable. Hassabis, who had once wished tech giants would move more slowly on AI deployment to ensure safety, was now the man pressing the accelerator.

Gemini generalist launch

While AlphaFold was winning prizes, the rest of Google DeepMind was fighting in the mud of the consumer market. The “Prosaic” reality of 2024 and 2025 has been defined by a relentless schedule of product releases, some revolutionary, others bizarre.

The flagship response to OpenAI was Gemini, a multimodal model family designed to power everything from Google Search to Android phones. Unlike the specialised AlphaFold, Gemini is a generalist, a jack of all trades designed to write emails, plan vacations, and code software. But the most peculiar skirmish in this war involved a model colloquially known as “Nano Banana” (Gemini 2.5 Flash Image).

In late 2025, this image generation tool went viral, not for curing cancer, but for a TikTok trend where users generated portraits of themselves across decades, from the 1880s to 2025. The model also gained notoriety for its ability to solve handwritten math homework, mimicking the user’s own handwriting style so perfectly that it sparked a debate about academic integrity. In one bizarre incident, an employee used it to generate a hyper-realistic image of an injured hand to fake a bike accident and get paid leave, prompting the viral tagline, “AI just broke HR verification.”

“Nano Banana” drives user engagement, locks people into the Google ecosystem, and demonstrates the “magic” of AI to the average consumer. The pricing models for these tools, ranging from free tiers to “Pro” subscriptions, are designed to monetise creativity at scale, a stark contrast to the open-science ethos of early DeepMind.

The threat to Google’s dominance intensified in October 2025 with the launch of ChatGPT Atlas, OpenAI’s AI-powered web browser. Atlas represents a paradigm shift. Instead of searching for links (Google’s model), users converse with the web. The browser features “Agent Mode,” where the AI can book flights, fill out forms, and summarise pages autonomously.

Atlas is a direct dagger at Chrome’s heart. If users stop searching and start “asking,” Google’s ad revenue, the lifeblood of Alphabet, evaporates. Hassabis’s team has responded with “Project Astra,” a universal AI assistant that can see and hear the world, integrated into Gemini Live.

AlphaFold solves mystery

Amidst the chaos of the chatbot wars, Hassabis delivered a reminder of why he started DeepMind in the first place. In 2024, the Nobel Committee recognised AlphaFold, DeepMind’s protein structure prediction system, with the Nobel Prize in Chemistry.

Proteins are the machinery of life. Their function is determined by their 3D shape, but predicting that shape from a string of amino acids is a problem of astronomical complexity. Levinthal’s paradox suggests it would take longer than the age of the universe to brute-force a solution.

AlphaFold 2, released in 2020, solved this. It predicted the structures of nearly all 200 million known proteins with atomic accuracy. The impact was immediate. Researchers used it to design malaria vaccines, understand antibiotic resistance, and develop plastic-eating enzymes.

For Hassabis, the Nobel was proof of his core thesis. He often said that the ultimate goal of AI is not just to create intelligent machines, but to understand intelligence itself.

AlphaFold was the perfect example of AI acting as a multiplier for human ingenuity, a “Hubble Telescope for biology.” In interviews following the award, Hassabis emphasised that scientific discovery was the true purpose of AI.

“I think we’re going to find… that some jobs get disrupted, but then new, more valuable, usually more interesting jobs get created,” he noted, framing AI as a tool for “radical abundance.”

However, the Nobel Prize also served as a shield. It gave Hassabis the political capital to push back against the complete commercialisation of his lab. It was a signal to the shareholders: “We are not just a chatbot factory. We are the Bell Labs of the 21st century.”

Transparency takes a hit

Training the next generation of AI models requires investment on a scale that rivals the “Manhattan Project.” This financial reality has escalated with the announcement of the “Stargate Project,” a massive $500 billion infrastructure initiative backed by OpenAI, SoftBank, Oracle, and the United States government.

This unprecedented capital injection into Google’s primary rival fundamentally alters the landscape. For Google to compete, it must match this investment dollar for dollar. Alphabet’s stock (GOOGL) has performed well, largely due to the perception that Gemini has stabilised the ship against the Microsoft-OpenAI alliance.

However, the transition from a high-margin search business to a high-cost AI compute business is risky. Every query answered by Gemini costs significantly more than a traditional Google search.

Demis Hassabis has had to make a devil’s bargain. To fund the “Profound” (AGI for science), he must win the “Prosaic” (commercial AI). “Commercial products fund science” is the unspoken mantra. The revenue from Google Cloud and Search pays for the TPUs that power “AlphaFold 3” and “AlphaProteo.” This reality has forced DeepMind to become less open.

The days of publishing every breakthrough in Nature immediately are gone. Now, technical reports are often withheld or redacted to prevent competitors like OpenAI and China’s DeepSeek from gaining an edge. The “Open” in OpenAI may be a misnomer, but Google DeepMind has also closed its doors.

Alchemist’s dilemma

Demis Hassabis stands at a crossroads. On one hand, he holds the Nobel Prize, a symbol of AI’s potential to elevate humanity. On the other hand, he holds the keys to the world’s most powerful ad-targeting engine, weaponised with generative AI.

The “Age of Paranoia,” fuelled by deepfakes and AI fraud, is rising alongside the “Age of Abundance” promised by AlphaFold. Hassabis’s challenge is to navigate this duality. He must ensure that the drive for profit does not corrupt the pursuit of discovery. The “Nano Banana” generated portraits and the “Atlas” browser wars are the noise of the present. They are the “Prosaic” tax that must be paid. But Hassabis’s eyes remain fixed on the horizon, on the “Profound.”

The young super-genius has come a long way from his early chess tournaments and video game development days. Hassabis has revolutionised how human beings think and act. His research in AI has also contributed to advancements in biology that would otherwise have taken another century.

No matter how things evolve from this point, Hassabis and his version of ethics will have a profound impact on how AI is used. He is the crusader fighting for the soul of Silicon Valley. Only time will tell whether science and human advancement will triumph against ads and corporate profits.

Demis Hassabis is one of the few individuals in history who simultaneously transformed science and business, which makes him both fascinating and concerning. On one hand, AlphaFold proves that AI can solve problems humans could not solve in decades. On the other hand, the commercial pressures of Google and the chatbot wars show that innovation is tied to profit.

Hassabis is balancing the desire to advance knowledge with the need to dominate markets. How he manages this will define whether AI truly serves humanity or becomes just another tool for corporate control. Right now, his choices are shaping the future of science, ethics, and the very way people interact with technology.

The post Demis Hassabis expands tech throne appeared first on International Finance.