AI of the Storm
A year ago AI was just a curiosity — now it’s a full-blown Category 5 hurricane tearing the roof off business, politics, and reality itself.
The initial blast hit with ChatGPT and NVIDIA’s trillion-dollar surge, but that was merely the outer band.
Headlines went quiet — like flags suddenly falling limp as we entered the eye of the storm — an eerie lull as the explosive cost of scaling foundational models stalled major breakthroughs. Yet beneath the surface, fierce competition raged on, labs trapped in a brutal game akin to opening yet another Thai restaurant in downtown Bangkok, forced to drip-feed partial updates — handing out their recipes and business plans one agonising ingredient at a time, each desperately clinging to relevance on AI’s ruthless Billboard charts.
Source (Midjourney & Grok)
This deceptive stillness concealed intensifying computational winds, businesses frantically racing to embed AI into… everything. The pace chaotic, the scale global, with no company, government, or job safe from the tempest.
Now comes the deadly back wall — the application revolution.
The next 18 months won’t be about gentle evolution. They’ll be extinction-level events for business models built on outdated assumptions about what machines can and cannot do. We are witnessing a fundamental reshaping of technology, economics, and geopolitics.
We stand in the eye of the AI hurricane.
Category 5 Models Make Landfall
Only a year ago, GPT-4 was the awe-inspiring peak of AI. Now a new class of models has crashed onto the scene like a storm surge. Anthropic’s Claude 3.7, billed as the first “hybrid reasoning” AI, can switch from near-instant answers to deep step-by-step logic on demand. It even writes production-ready code via an agentic toolchain called Claude Code, essentially automating chunks of software engineering from a command line. The “Midjourney Moment” for software engineering is about to arrive.
Meanwhile, Elon Musk’s xAI has leveled up Grok to version 3, a model with a sarcastic streak and multimodal chops, and Google’s DeepMind/Brain team launched Gemini 2.0 to one-up them all in multilingual, image, and audio tasks.
Then China entered the room.
OpenAI’s Business Model in Turbulence
Not long ago, OpenAI looked invincible — the tollbooth operator on the highway to humanity’s collective intelligence, perfectly positioned to charge every passing thought as they were poised to become the API for all intelligence.
Pay them a few cents and get magic out of the box. It was a great story while it lasted. Today that story is dead. The commoditisation of large language models (LLMs) has shredded OpenAI’s once-impenetrable moat. As I predicted early on, there was never a sustainable competitive advantage in simply serving up a generic AI model via API. The core technology was too portable, and there’s literally zero switching costs. Indeed, the moment a better model came along, people switched en masse — there’s nothing to lock you in other than maybe a chat history.
We saw this clearly with Anthropic’s Claude. One day I’m using GPT-4 for everything; the next, Claude 2 lands, and suddenly it’s dramatically better at writing what I need. I dump ChatGPT overnight. No one in the real world cares that OpenAI sunk $100 million in compute to train their model — if a rival’s output is better or cheaper, we’re gone. This basic truth has played out again and again: from GPT-3 to GPT-3.5 to Claude to GPT-4, and now to a swarm of new models (Claude 3.5, 3.7, GPT-4.5 and beyond). Each time a new contender delivered an incremental improvement or specialised skill, users migrated without hesitation. The era of a single dominant AI API has ended almost as quickly as it began.
Soon enough, users tire of paying $20 per month for models that become obsolete the moment something smarter arrives. After all, who wants to keep paying for yesterday’s inferior assistant when today’s sharper, faster, more capable version is already here? It’s not subscription fatigue — it’s redundancy fatigue.
It’s déjà vu watching OpenAI flail around trying to find a business model — first $20 a month for ChatGPT Plus, then tossing out $200-a-month “enterprise-grade” bait, and recently floating the eye-watering $2,000 and even $20,000-a-month “agent architecture.” Remember when they tried a marketplace for those so-called “GPTs,” miniature AI bots promising to handle everything but ending up mastering nothing, quietly vanishing into the great digital beyond? Well, the word is they’re back at it again, warming up yet another marketplace reboot.
At this point, maybe they should just call it Slopify.
OpenAI’s leadership learned this the hard way. They tried to erect an ecosystem lock-in — app stores, plugin platforms, proprietary tooling — to cement their monopoly. It isn’t working. Why use a rigid API with usage caps, content filters, and latency issues when you can download an open-source model and run it yourself for free?
By late 2024, some companies started quietly shifting gears: instead of relying solely on OpenAI’s API, many began experimenting internally with smaller, specialised models or increasingly turned to the vibrant, rapidly advancing open-source communities.
OpenAI slashed its API prices, but even free might not be cheap enough if a competitor is even slightly better. The commoditisation trap has snapped shut, and OpenAI is caught in it.
As I said on MacroVoices last year, charging a few cents per API call was never a defensible business model — “the end reality is, there’s no actual business model here with these foundational models”. That reality is now impossible to ignore.
Open-Source Dominance and the Collapse of the Moat
If 2023 was the year proprietary models stunned the world, 2024 was the year open-source took the crown. The radical proliferation of open-source LLMs and tools has utterly annihilated the “moat” that Big Tech and well-funded startups thought they had. A leaked Google memo from last year said it best: “we have no moat, and neither does OpenAI,” because a “third faction” — the open-source community — is “faster, more customizable, more private, and pound-for-pound more capable” than the tech giants. That prophetic memo has been borne out in spades.
Facebook (Meta) tried to lead the way by releasing its LLaMA models, deliberately dumping at least part of its crown jewels out into the wild to kick the chair out from under OpenAI — though calling it “open source” was always a stretch. It was corporate kamikaze, and even with the restrictions, it worked. Within weeks, LLaMA’s weights inevitably leaked, and hobbyists quickly fine-tuned versions rivaling GPT-3. By the time Meta rolled out LLaMA 2 under its controversial, not-quite-open licence (much to the irritation of the open-source purists), the community was already iterating faster than many well-funded labs. Suddenly, a global army of researchers and hackers — unfettered by NDAs, profit targets, or strict licences — managed to replicate years of secretive corporate R&D in mere months. The pace of improvement in community-driven models turned out to be faster than anything closed shops could handle. OpenAI’s and Google’s worst nightmare materialised: the bleeding edge of AI innovation no longer belonged exclusively to them; it was now happening out in the wild, licence be damned.
Look at what happened in late 2024: a Chinese startup called DeepSeek came out of nowhere and triggered a shockwave across the industry. DeepSeek’s team, working with modest resources, trained a reasoning model (“R1”) that outperformed many Western models built at 100x the cost. How? By leveraging every open-source insight available, innovating on efficiency, and presumably a healthy disregard for Silicon Valley orthodoxy.
Their R1 model ran on comparatively weak hardware yet delivered competitive results — proof positive that the AI crown jewels are now knowledge, technique, and execution, not access to infinite GPU farms. When news broke that DeepSeek’s cut-price model was matching the giants, it erased over a trillion dollars in market cap from frightened investors. I mean, 160 unknown engineers working on a side project in Hangzhou optimised their way into the same league as OpenAI’s GPT-4 — and the market rightly panicked over the trillions of dollars being thrown at western AI.
The implications are profound. First, it validated the open approach: DeepSeek didn’t benefit from secret proprietary datasets or magic algorithms unavailable to others. They stood on the shoulders of open-source and pushed new optimisations (we’ll talk about their hardware tricks shortly). Don’t forget that Ilya Sutskever, Co-founder and former Chief Scientist of OpenAI said originally that 90% of AI is explained in 40 papers that are publicly available.
Second, it obliterated any remaining argument that the incumbents’ “scale advantage” would keep them safe. If anything, the opposite is true — being big has become a liability. While OpenAI, Microsoft, and Google are dragging their enormous cost structures and bureaucracy, smaller players are moving at hyperspeed. As one Google engineer lamented, open-source models are iterating faster and more privately than the big labs can. The flood of innovation is now coming from the periphery.
Even OpenAI’s latest offering, GPT-4.5, truly is impressive when used with Deep Research, it’s like having a mid-level research assistant at your beck and call. The problem for OpenAI is that it’s decidedly mid, in my experience of using it, when “Deep Research” isn’t turned on, it can’t write code well. The window of advantage is rapidly closing.
DeepSeek’s upcoming model, R2, is already rumoured to score around 90% on ARC-AGI, a benchmark explicitly designed to measure AI against human-level intelligence. If that holds true, OpenAI’s incremental improvements won’t just feel modest — they’ll feel irrelevant. GPT-4.5 is incredibly expensive at $75/million input and $150/million output tokens, and that’s without “Deep Research”, which currently isn’t available via API. The code for R2 is likely to be free, and running on your own hardware, especially Chinese hardware, cheap.
We have reached a point where any novel model architecture or training method gets reproduced in an open variant by the community in a matter of months, if not weeks.
It’s hard to overstate what this means: the moat isn’t just draining; it’s evaporating. The formula for building a capable AI model is public knowledge. Yes, it still requires skill (and some capital), but not the kind of billion-dollar war chest and years of work that people assumed in 2022. We’ve seen university groups, independent researchers, and small startups produce models that would have been considered state-of-the-art not long ago.
At the centre of this AI storm sits OpenAI — once the undisputed pioneer, now scrambling to survive the onslaught. Its flagship ChatGPT still dazzles, but the company is haemorrhaging cash. OpenAI reportedly burned $7–8 billion in 2024 alone and forecasts a $20 billion burn rate in 2027, chasing ever-bigger models. As AI investor Kai-Fu Lee bluntly put it, “the issue isn’t whose model is 1% better… the issue is: Is OpenAI’s business model even sustainable?”.
The harsh truth is that foundation models have become commodities. When a Chinese open project like DeepSeek can replicate 90% of GPT-4’s performance at 2–10% of the cost, it undercuts the entire OpenAI model of pricey API access. It’s hard to charge premium rates for a product a well-funded competitor offers essentially for free.
China, of course, also holds an unshakable competitive advantage when it comes to training AI models:
They don’t give a rat’s arse about copyright.
While American companies are tiptoeing around licensing agreements, paying an army of lawyers to agonise over whether GPT can recite two lines from Harry Potter without getting sued, China is cheerfully scraping the entire internet — including your tweets, emails, and probably your grandma’s secret recipes — and pumping it straight into their neural networks. Forget data lakes; China’s running a data ocean. It’s like entering a boxing match with one arm tied behind your back while your opponent brings a chainsaw. Good luck winning that fight.
I recently talked about this with an Anthropic engineer. Nonchalantly they said “Why do you think we have offices in Japan and Korea now?”.
I’ve never liked first mover advantage- the first mover has to burn huge amounts of energy boiling the ocean to change consumer behaviour. Last mover advantage, on the other hand, gets to coast in the slipstream, riding comfortably on the coat tails of whoever’s out in front.
I wouldn’t be surprised if next month’s headlines feature fifty unknown engineers in a Bangalore basement launching an open-source model that makes ChatGPT look like a pocket calculator — a delicious rebuke to Sam Altman’s patronising dismissal of India’s AI ambitions as “totally hopeless”.
Remember his visit? When asked about India’s role in AI’s future, he essentially patted the country on its head and suggested it stick to being a good consumer rather than a creator.
Talk about a statement ageing like milk left in the Rajasthani sun.
We’ve already seen that breakthrough models don’t require billion-dollar budgets or Stanford PhDs — just ingenuity, determination, and perhaps a dash of righteous indignation. The next AI “Thai restaurant” could easily spring up from an unexpected corner, recipes in hand, ready to redefine the menu altogether.
The open-source dominance isn’t just in text generation. Image models (Stable Diffusion, etc.) proved it first in 2022. By now we have open models for coding assistance, speech synthesis, even multimodal vision-language tasks. The frontier of AI has fragmented into countless community-driven projects. No wonder OpenAI is scrambling — Altman went from declaring open-source models “dumb” a year ago to calling DeepSeek “impressive”.
Life comes at you fast.
For all OpenAI’s lofty talk of altruistic AI, David Sacks’ criticism nails it: they’ve “gone from nonprofit philanthropy to piranha, for-profit company”. In just a few years, Sam Altman’s outfit went from idealistic research lab to cutting backroom deals giving Microsoft huge financial and IP entitlements — think exclusive licences to OpenAI’s crown-jewel tech (Microsoft alone gets unique access to GPT-3’s underlying code), priority integration via Azure (Microsoft is OpenAI’s exclusive cloud fiefdom), and a staggering 75% cut of profits until Microsoft recoups its billions of investment back (translating to a capped return of up to 100×, meaning a $1 billion stake could fetch $100 billion).
All these sweetheart terms might be tolerable if OpenAI had an impregnable lead, but it doesn’t.
Meanwhile, Altman’s leadership is under fire from all sides: he picked a fight with Elon, only to find that Elon bought the most powerful political weapon in the world, Twitter, which he’s leveraged to now become the US Secretary of Spending via DOGE. On top of that Elon’s buddy David Sacks (who made the piranha comment) is now the AI Czar and Elon Musk is dragging OpenAI to court over its mission drift to for-profit status, and even OpenAI’s own board briefly ousted Altman last year for “no confidence” in his leadership before an investor revolt forced his reinstatement. Plus, 8 of the 11 original co-founders have already walked.
What a mess.
If I were a betting man, I’d reckon that Sam might be close to jumping ship and starting a new AI company free of this quagmire before his Thielian Zero to One founder story arc inevitably swings from hero to villain.
If that happens, Microsoft will do what it always does, ‘embrace and extend’ itself over the carcass.
The Hardware War: When it Rains Models & Sanctions, China Pours Alternatives
Nvidia today is sitting pretty at the eye of the AI hurricane — still the undisputed king of silicon for training giant models, printing money as fast as TSMC can etch transistors. Its latest GPUs (think H100s priced like sports cars) fly off the shelves, and its stock charts only seem to know one direction. But even as Jensen Huang basks in the sunshine of record profits, storm clouds are gathering on the horizon. The cracks in Nvidia’s monopoly are subtle but multiplying: the rapid obsolescence of each generation of hardware, the looming threat of vertical integration by its biggest customers, and the geopolitical rift splintering the global chip supply. Nvidia’s still on top, but the ground beneath it is starting to shift.
Let’s start with the hardware treadmill. In AI land, today’s cutting-edge chip becomes yesterday’s news fast. Nvidia’s business model loves this — persuade everyone that you need the latest GPU or you’ll fall behind. But that cadence of obsolescence is a double-edged sword. Each new generation renders billions of dollars of last-gen hardware somewhat redundant, straining the budgets of even the richest hyperscalers. It’s like a never-ending arms race where you upgrade or die. The result? Even Nvidia’s best friends (the Googles, Amazons, Microsofts of the world) are getting fatigue and looking for an exit ramp- 46% of Nvidia’s revenue comes from just four companies. Vertical integration is their escape route. Why buy $10 billion worth of GPUs every year if you can invest a fraction of that to design your own AI chips tailored to your needs?
Google has been quietly doing this for years with its TPUs — tensor processors that handle everything from Search to YouTube to DeepMind’s workloads without a single Nvidia chip in sight. Amazon’s AWS, not to be outdone, rolled out Trainium for training and Inferentia for inference, silicon designed in-house to keep those AWS bills in check. Microsoft is reportedly cooking up its own AI accelerator (Project Athena) to lessen its dependence on Nvidia in the long run. In short, the largest buyers of Nvidia hardware are developing their own alternatives.
That’s a hairline crack in Nvidia’s dominance that could widen fast: when your biggest customers become competitors, the rainmaker can quickly find themselves in the rain.
Then there’s the geopolitical deluge. The U.S. decided to play weatherman and imposed heavy sanctions on exporting top-tier AI chips to China, aiming to slow down China’s AI progress. Nvidia, ever the savvy, responded by creating “China-friendly” versions of its GPUs (read: slightly neutered H100s like the H800 and new H20 with capped interconnects) to comply with the rules and still sell something to a market that recently made up nearly 20% of its revenue.
In 2019, Huawei released its top-of-the-line Ascend 910 chips fabricated on TSMC’s 7nm process node. In response, the U.S. government put Huawei on its Entity List in 2020 sanctioning it. But Washington’s move also had an unintended effect: it galvanised China to double down on its own semiconductor efforts.
When it rains models and sanctions, China pours alternatives.
Enter Huawei’s Ascend silicon.
A few years ago, Huawei’s AI chips were seen as curiosities — good for a patriotic headline, not a serious threat to Nvidia.
Not anymore. Huawei openly wants to become the “Chinese Nvidia” and with recent leaps, they’re not far off.
Huawei is ramping up production of its Ascend 910 series, even under sanctions. These aren’t vapourware like Apple Intelligence; they’re real hardware deployed at scale.
Huawei had to redesign its Virtuvian chiplet to make it at SMIC, which used its N+1 technology (1st Generation 7nm-class process) to reengineer the 910B avoiding those sanctions.
The second-generation Ascend 910B chips have higher maximum performance, taking the 910 which ran at 320TFLOPS to 910B at 400TFLOPS. It has a higher clock speed but fewer active AI cores, one additional vector unit in each AI core, and a newer memory type with higher bandwidth and capacity. Huawei reduced the number of active AI cores between the 910 and 910B series — likely either due to poor yields or limited capacity on SMIC’s 7nm fabrication process.
50% of the increase is due to the increased clock speed. An additional 25% can be attributed to the additional vector unit in each AI core. The remaining 25% of the performance increase between the two generations of chips appears to result from a change in how Huawei calculates peak performance.
The Ascend 910C is simple two 910Bs on a chip and runs at 800TFLOPS. It’s built on a Chinese 7nm process (SMIC’s N+2 node) under a sanctions cloud — no EUV lithography, many skeptics — and yet it delivers about 60% of an Nvidia H100’s performance on inference tasks. That’s right: a sanctioned 2019-era design, upgraded and re-packaged, now holding its own in the domain of running large AI models.
Yields have been low and it’s costing them a fortune thanks to sanctions forcing them to use less advanced fabs and fewer engineers for the management and maintenance of chip tools, but they’re making rapid progress. According to reports, yield on the 910C has already improved from 20% to around 40%, and yield improvements have now made Ascend profitable. In short: China is throwing brute force, money, and top engineering talent at breaking the NVIDIA bottleneck. They will get there. If not this generation, then by the next. The U.S. had a fleeting window of unquestioned dominance in AI hardware — that window is closing, fast.
Huawei needed at least two years to redesign and domestically fabricate the second-generation Ascend 910B chips, with only marginal improvements.
In 2024, Huawei built 200000 Ascend 910B processors and no 910Cs.
This year, they plan to develop 100000 910C and 300000 910Bs.
On top of that, TSMC reportedly manufactured more than 2 million Ascend 910B logic dies illegally through shell companies and that all of these are now with Huawei, enough to make 1 million 910Cs.
For training giant models, Nvidia still has a comfortable lead (software ecosystem, maturity, reliability — decades of head start don’t vanish overnight). But for inference — the workhorse of AI deployment — Huawei has effectively closed much of the gap. And here’s the kicker: Huawei is flooding the market with these things. Despite the expected teething issues (yield headaches on a cutting-edge process, the complexities of chiplet packaging with 50+ billion transistors, etc.), they are turning up production like there’s no tomorrow.
The impact is being felt in pricing. Chinese cloud providers, under gentle nudges from Beijing to “buy local,” have embraced the Ascend chips to power their AI services. The result: inference costs that Western firms simply can’t match right now.
Imagine calling an Uber and getting charged 5 cents, because that’s basically what’s happening for AI inference in China.
DeepSeek R1 is offered on cloud platforms like SiliconFlow for the jaw-dropping price of ¥1 per million input tokens — roughly $0.15. For output tokens, it’s about ¥4 per million (~$0.60). Even combined, that’s well under a dollar per million tokens processed.
For comparison, OpenAI’s vaunted GPT-4, with all its Microsoft Azure-backed infrastructure, charges about $30 or more per million tokens when you tally up prompt and completion pricing. We’re talking a 200× price difference. Yes, the models aren’t identical in capability, but the economic moat that outfits like OpenAI/Microsoft enjoyed — the assumption that only they have the scale to serve millions of users — is being blown out of the water.
When a Chinese cloud can serve a GPT-4 class model for pennies on the dollar, the whole narrative of “AI is super expensive, thus premium pricing” starts to crumble. OpenAI finds itself in a price war it never wanted, against an opponent (the Chinese AI ecosystem) that isn’t playing by Wall Street’s playbook of fat margins. It’s hard to sell $0.03 per 1,000 token API calls when a competitor is effectively offering it for $0.00015 via subsidy or sheer cost advantage. In a matter of months, the price floor for AI services has dropped through the basement, and Nvidia’s hardware hegemony is partly why — or rather, the cracking of that hegemony.
Now, about that DeepSeek R1 model and why every AI investor spat out their coffee in late January. If the rise of Huawei hardware was one front in this war, the assault on brute-force thinking is another.
For years, the playbook in AI has been straightforward: got a bigger model to train? Throw more GPUs at it (ideally Nvidia’s, of course). Scale, scale, scale — and damn the efficiency. DeepSeek threw a delightful wrench into that approach. Faced with only “bargain-basement” H800 GPUs (the toned-down version of H100 that Nvidia can legally ship to China), the engineers at DeepSeek decided to squeeze blood from a stone. They went low-level — we’re talking below CUDA, right into Nvidia’s PTX assembly instructions. If you’d like to deep dive into how it works, check out this article.
It’s the kind of hardcore optimisation work that most teams don’t even contemplate because it’s insanely difficult and ties your software to specific hardware in a non-portable way. But desperate times in China called for desperate measures. And it worked — spectacularly. By reconfiguring how their code ran on the H800’s cores, optimising memory access patterns, and offloading certain tasks to portions of the GPU that normally sit idle, DeepSeek achieved an astonishing 10× training efficiency gain over the baseline. Let that sink in: not 10% — ten-fold. They effectively turned each H800 into something more like an H100 on steroids. With just 2,048 GPUs and a rag tag team working on a side project, DeepSeek trained a 671-billion-parameter mixture-of-experts model (DeepSeek R1) in about two months.
For comparison, Meta needed around 2,000 top-of-the-line Nvidia GPUs (A100s) to train its 65-billion-parameter LLaMA model over a similar timeframe. DeepSeek was over an order of magnitude more efficient in hardware usage for a model 10× larger. This is the part in the movie where the music stops and everyone looks around, muttering “Wait, that’s possible?” It wasn’t supposed to be — but they did it anyway.
The repercussions of DeepSeek’s achievement hit Nvidia where it hurts: its market cap. The moment investors realised that a hungry startup had essentially made do with fewer Nvidia chips to accomplish a Herculean AI task, the narrative of infinite GPU demand took a hit. Nvidia’s stock, which had been on a parabola rocketing up on every whisper of AI hype, suddenly felt gravity again. In the span of days, Nvidia has shed $866 billion.
Now, stock swings are often overreactions, but symbolically it was clear: the era of winning just by having the most brute-force compute might be nearing its end. Optimisation is the new scaling. When a single clever trick or algorithmic improvement can save you 90% of your compute budget, you pay attention. And if you ignore it, your competitors won’t. Even a modest 2× efficiency gain — heck, even 20% — at the scale of global AI deployment is a game-changer. That’s the difference between needing five data centers versus four, or spending $100 million on GPUs versus $80 million. Those sums make or break quarterly earnings. DeepSeek proved that brute force can be outsmarted, and in doing so, it chipped away at Nvidia’s aura of indispensability.
Add all this up and what do we see? A brewing storm of fragmentation in the AI hardware landscape. Just a year ago, one could be forgiven for thinking Nvidia had won the war — their GPUs were in virtually every AI lab and data center, their software stack CUDA was the lingua franca of AI, and challengers were dismissed as academic footnotes or niche players. How quickly things change when the stakes are this high. We’re now hurtling toward a decentralised, multi-polar era for AI hardware. In one corner, Nvidia’s GPUs aren’t going anywhere — they will continue to be the gold standard for many tasks, and Jensen will undoubtedly come back swinging with even more powerful silicon (Blackwell and beyond) and improved software.
But the monopoly moment is fading. In another corner, we have China Inc., led by Huawei (and surely others to follow), forging an independent path — one that prioritises sufficiency and scale over bleeding-edge supremacy. They may lag slightly in absolute performance, but they’ll make up for it in volume and cost, and they’re free from Western chokeholds. In yet another corner, the cloud hyperscalers and a slew of startups are building a Noah’s Ark of novel chips: TPUs, IPUs, NPUs, analog chips, photonic chips — you name it — each optimised for specific workloads or efficiency targets.
It’s as if the AI compute universe has split into multiple parallel timelines, all unfolding everything everywhere all at once.
This fragmentation isn’t a sign of slowing innovation; it’s a sign of commoditisation and democratisation. Compute is becoming a commodity in the sense that everyone will have access to serious AI horsepower, one way or another. When multiple countries and companies can each mass-produce “good enough” chips, no single player can corner the market. Prices will fall; margins will tighten. (Nvidia won’t be selling $40,000 boards into eternity — competition will see to that.) At the same time, the real differentiator will shift from raw compute to how cleverly that compute is used. This is where optimisation as the new paradigm comes in. The lesson of DeepSeek will echo: instead of just throwing money and silicon at the problem, throw brains at it. Make the algorithms leaner, the software tighter, the hardware more specialised. It’s a return, in some ways, to the early days of computing — when efficiency mattered because hardware was limited. Now hardware is abundant but still expensive at scale, so efficiency is king again.
The global AI hardware landscape by the end of this stormy decade will look nothing like the monolithic one we had at the start. It will be geopolitically fragmented — a U.S.-led sphere, a China-led sphere, and likely others joining the fray (Europe, India, whoever stakes a claim). It will be technologically diverse — GPUs sharing the stage with ASICs, FPGAs, and whatever wild new architectures come out of left field. And it will be far more resilient — no single point of failure or control, no single chokehold that can be exploited to stall progress. Nvidia’s GPUs will still be immensely important (you’d be foolish to bet against them outright), but they’ll be one among many options rather than the only game in town. The era of one-size-fits-all, general-purpose GPU dominance is yielding to a patchwork of specialised, optimised compute scattered across the globe.
In the final analysis, the hardware war in AI won’t crown a solitary winner. Instead, it’s raining on the once-unified parade and giving rise to a lush, if messy, rainforest of alternatives. Each ecosystem — Nvidia, Huawei, Google, open-source efforts — is like a different species evolving to fill its niche. And ironically, this diversity will drive AI forward faster than a single-player monopoly ever could. The storm has well and truly arrived, and it’s washing away assumptions left and right.
For those of us watching from the sidelines (investors, researchers, policy makers), one thing is clear: the future of AI will be built on a globally distributed web of silicon, and the smartest approach to wielding it — not just the biggest chips — will determine who leads the next chapter of this story.
The 10× training efficiency gain that DeepSeek achieved, where a 671-billion-parameter model with just 2,048 of NVIDIA’s H800 GPUs took just two months compares to something Meta or Google would normally throw 20,000 GPUs at.
DeepSeek effectively neutralised NVIDIA’s hardware advantage by using brains over brawn. And in doing so, they proved that the global GPU shortage and U.S. export bans can be at least partly circumvented with ingenuity and elbow grease.
If one upstart can do that, others will follow. We’re going to see a wave of “PTX-level” optimisation as every major AI player in China (and likely Russia, and even hungry startups in the West) tries to maximise performance on whatever hardware they can get. This erodes NVIDIA’s de facto monopoly further, and it accelerates the trend of AI compute decentralising.
Already, alternative AI chip startups (from Cerebras in the US to Biren in China) smell blood in the water. By 2026, the hardware landscape for AI could look very different — a mix of NVIDIA, domestic Chinese AI accelerators, and exotic new architectures optimised for local running of models. In any case, NVIDIA won’t have the chokehold it had at the peak of the boom. The hardware war is breaking the world into camps — and ironically, it’s also pushing everyone towards efficiency over brute force.
One final twist: while the world’s appetite for inference compute will undoubtedly explode — driven by a tidal wave of AI applications we’ve barely even deployed at scale yet — the critical question is where that compute will live. As inference migrates from centralised cloud fortresses out to the edge (think smartphones, vehicles, factories, and sensors everywhere), many of today’s giant data-center builds could suddenly look a lot less strategic. We’re currently witnessing a classic boom-and-bust scenario: hyperscalers racing to fill entire football fields with racks of GPUs, potentially overshooting demand if the centre of gravity shifts toward decentralised inference at the edge.
Still, it’s worth remembering what I call Scroggie’s Law: Compute inevitably expands to fill the available data centers, no matter how many you build.
The Rise of AI Agents: Embedded Intelligence at the Edge
For all the focus on grand AI models in the cloud, perhaps the more profound shift happening is the migration of intelligence outwards — into real-time agents, devices, and the edge. We’re moving from an era where AI meant “call an API to get an answer” to an era where AI is embedded everywhere you need it, often running locally, autonomously, and continuously. This has massive implications: it upends the cloud-centralised dominance of AI, changes how software is built, and promises to boost productivity in countless niches that big models-in-the-cloud could never efficiently serve.
In 2023, the buzz was all about ChatGPT in your browser. In 2025, the buzz is about AI copilots in everything: your word processor, your car, your fridge, your company’s customer support workflow. These are AI agents — specialised, task-driven AIs that don’t just respond passively to prompts, but actively take initiative, observe, plan, and execute goals in real time.
Early experiments like AutoGPT and BabyAGI were clunky proof-of-concepts, but they lit the spark. Now we have evolved versions deployed in production.
At my own company, Freelancer, we built an AI agent framework to handle tier-1 customer support and sales inquiries. These agents can interact with customers in natural language, answer questions, troubleshoot issues, and even upsell our services — all without a human in the loop until a certain point. We also do this for others- every small business in the world in the next two years will have AI answer the phones, make a booking, take a credit card and process an order- check out freelancer.com/ai for some demos.
And here’s an intriguing finding: rather than simply replacing humans, these agents augmented our team, taking on the drudge work and creating new tasks for humans when they bump into limits and need to escalate, thus actually increasing our net employment.
It’s counterintuitive but makes sense — the AI handles 1,000 simple tickets it would never have been economical for a human to address individually, but in doing so it uncovers 100 cases that need higher-level human intervention (thus generating work that wouldn’t have existed). We’ve essentially hired a team of tireless, ultra-fast juniors that escalate to the (human) managers when needed. And this is just one example in customer support.
On top of that, we have net additional product teams now working on the framework itself- so at least in our case and for now, like all technologies in the past they’ve created more jobs than they’ve destroyed.
Now extend this pattern across industries. Every business process that involves routine communication or data processing is ripe for AI agent automation. We’re talking sales calls, tech support, HR onboarding, internal helpdesks, research analysis, personal assistants — the works. In real estate, as I mentioned in a past Macrovoices interview, imagine an AI agent handling all the annoying rental property management calls — tenants calling about a broken tap, scheduling a repair, updating the landlord.
An AI could triage and respond to all of that, looping in a human only for the actual physical repair. Or consider medicine: AI “transcription agents” already listen in on patient visits and draft clinical notes; the next step is an AI agent that can do initial patient intake, ask questions about symptoms, and maybe even suggest a preliminary diagnosis for the doctor to review. That’s not sci-fi — components of it exist now. If the government would let ChatGPT write a script, it would probably free up half of the GPs in the country for higher value work.
Crucially, many of these agent applications demand real-time or on-site AI. You might not want a cloud server handling a sensitive business phone call if a local AI can do it with no latency and full privacy. Also, as I mentioned in AI Know What You Did Last Summer, as AI features get better (and creepier), I think SaaS is about to have its Emperor has no Clothes moment.
I think that you’re going to increasingly see a lot of consumers and companies saying “I don’t want my data on the Internet, I don’t want the AI to suck it down, I don’t want the AI knowing about my user base, I don’t want them knowing about my business model, I don’t want my research to be instantly commercialised”.
You might see the Internet going dark in a very big way.
I mean just go look at your Gmail now and think to yourself..
Google’s AI already knows everything that’s in there.. Everything.
They’re lying to you that they don’t look at that data, they serve contextual ads to you based off of it.
Hey Google, please tell me the best way to compete against my company?
The drive for embedded inference — running AI models on local devices (phones, laptops, on-prem servers, cars) — has kicked into high gear. The same way the first phase of the personal computer revolution moved computing from central mainframes to individual desktops, we’re now moving AI from central data centers to personal and edge devices.
In 2024, Qualcomm demonstrated running large language models on smartphones; by 2025, it’s becoming common for consumer devices to ship with on-device AI capabilities that would have required an internet connection before. Apple’s Neural Engine, Google’s Tensor chips — they’re all being purposed to run models directly on your gadget. And open-source made this possible: thanks to model compression techniques (quantisation, distillation), people can squeeze what was a 13 billion parameter model in 2022 down to something that runs acceptably on a $1,000 laptop today.
This end of cloud AI dominance is a nightmare for certain big tech business models. If the centre of gravity shifts to the edge, cloud AI providers may find their growth curtailed. Why pay per token or per month for an API if you can own the model that runs locally with no ongoing cost?
We’re already seeing early signs: savvy companies are taking open models like Llama 2, fine-tuning them on their proprietary data, and running them internally — no OpenAI or Azure needed. The cost to do so has plummeted. It’s now feasible for a mid-sized business to have its own GPT-esque model in house, tailored to its needs (and far less likely to spill secrets to outsiders). That undercuts the “AI-as-a-service” (ASS) paradigm significantly. I mean, how many of you have sometimes had a second thought before you typed something into ChatGPT?
I even have some friends that make sure they’re really nice to ChatGPT every time they type into it so they remember them when Skynet becomes self aware.
AI in the edge also heralds a renaissance in software development. Instead of thinking of “AI” as this separate thing, developers are beginning to treat it as just another library or runtime they include in applications. The next generation of apps will have AI baked in at the core. Imagine opening Excel and not just seeing static formulas, but an AI agent in the sheet that can analyse your data, build a pivot table, generate a forecast, and even highlight anomalies — all running on your CPU or GPU. Microsoft is already pushing toward this with Office 365 Copilot, but right now it calls out to the cloud. Give it a year or two and your PC will likely handle a lot of that inference itself, especially for smaller-scale tasks.
One interesting implication here is the rise of vertical AI agents. As generalised models commoditise, value shifts to how you use them in context. An AI agent that is an expert in, say, supply chain logistics, or medical billing, or legal contract analysis — those are vertical solutions built on top of possibly commodity models but enhanced with domain data and process integration. The money to be made is in deploying AI to solve specific pain points, not just in having the fanciest model. I’ve said it before: the real money is in applying AI across every industry, the same way in the dotcom era the money was in putting every business online. We’re seeing that play out. Every industry will have its bespoke AI assistants and agents and the Internet will go dark for data sets that are both new and public (You can’t just scrape ArtStation for training data anymore can you?).
Many of those vertical agents will run on-premises or on specialised devices, for reasons of reliability, latency, and confidentiality.
Consider also the autonomy aspect. These agents aren’t just Q&A bots; they can act. We’re effectively instilling a level of decision-making autonomy in software that used to always be scripted by humans. This is powerful, but you better be damn sure the agent is doing what you intend. If there’s one area I urge caution, it’s unleashing autonomous agents without guardrails. We learned from the first wave (some AutoGPT users found the agents looping endlessly or doing stupid things). The next wave, hopefully, we’re building with more robust objectives and safety checks against doing something stupid (like an AI Travel Agent throwing out excessive discounts, or doing people’s maths homework on request).
Autonomous AI agents can execute trades, send emails, interface with databases, control IoT devices. That introduces a whole new set of opportunities and risks. A trading firm might have an AI agent monitoring market conditions and executing strategies faster than any human. A factory could have AI control systems that dynamically adjust operations on the fly. A security camera will know everything about a scene and do reverse lookups on everyone’s identity from their face, gait and other features.
Nonetheless, the genie is out of the bottle. Autonomous AI agents are here to stay, and they will only grow more capable. Real-world effectors (like robots or software privileges) combined with AI brains create a feedback loop: the AI can try something, observe the outcome, and learn — a rudimentary form of iteration towards goals. It’s not self-aware Skynet, but it doesn’t have to be to revolutionise workflows. One well-designed customer service agent that continuously learns from successful and failed interactions could outperform an entire outsourced call centre department in a year’s time.
Summing up: we are decentralising and specialising AI. The giant cloud model that does everything okay is being supplanted by swarms of smaller, more focused agents that live closer to where the action is. This shift is akin to moving from mainframes to PCs, or from monolithic apps to microservices. It changes who captures value (expect edge device makers and savvy integrators to benefit, while cloud usage growth may slow). It also challenges the notion that “AI = big data center.” In a couple of years, that may sound as outdated as “internet = AOL.”
Regulation GPT: Government Panic Theatre
If you thought scaling neural nets and hardware wars were chaotic, wait until you see what governments are cooking up. The U.S. bans chips to China; China bans export of certain AI software; the EU debates forcing “GAI” models to get a government licence; and so on. Welcome to the global circus: AI nationalism meets bureaucratic bumbling.
Europe is out front, racing ahead with its monumental EU AI Act, driven by their tried-and-true ‘precautionary principle’ — which essentially means regulating first, innovating second. It’s classic EU: outright bans on certain uses and heavy compliance demands for “high-risk” systems, massive compliance costs, sweeping restrictions, endless paperwork, and fines so big they could bankrupt small countries (up to €35 million or 7% of global revenue). Europe might just regulate itself out of relevance, turning their continent into a digital theme park, while the rest of the world builds the actual AI future.
The US, by contrast, is sticking to its “move fast and break things” mantra, issuing AI rules via executive orders, guidelines, and voluntary frameworks. It’s regulation-lite, American-style — leaning on Silicon Valley’s self-regulation while nervously wanting a kill-switch, or at least an off switch, for AI — they just haven’t figured out how to install it without blowing up the economy. Biden’s executive order basically asks tech giants nicely to play fair, share their homework, and watermark their deepfakes. Meanwhile, Congress debates AI regulation with all the urgency of a glacier — hoping nothing catastrophic happens on their watch.
China’s strategy couldn’t be clearer: innovate fast, censor even faster. Beijing’s new rules force AI to uphold “core socialist values” — no surprises there. Every AI-generated byte is monitored, filtered, and obediently stamped “Made in China.” And unlike their Western counterparts, they don’t even pretend they’re not watching. If a user somehow coaxes the AI into producing banned material, the provider must immediately halt generation, delete the output, and report the incident to authorities. This isn’t just regulation, it’s AI with Chinese characteristics — efficient, innovative, and absolutely under control.
This three-way tug-of-war — Europe’s cautious nanny-state, America’s laissez-faire innovation, and China’s authoritarian alignment — is tearing the AI ecosystem apart. Companies must now geo-fence their models, maintaining separate China-only versions of their models with censorship built-in, and deal with U.S. export controls on their hardware or even their models (imagine needing a Commerce Department licence to share a weights file!).
AI startups face a compliance nightmare, and global platforms have become schizophrenic, changing personality and functionality based on your passport. We’re heading towards a fragmented AI landscape, split between the EU’s walled garden, America’s open-but-risky playground, and China’s state-controlled digital fortress.
As the Economist quipped at the dawn of 2024, “welcome to the era of AI nationalism,” where “sovereigns the world over are racing to control their technological destinies”.
One wildcard in all this: the open-source AI rebels. They’ve democratised foundational models, blowing holes in the proprietary moats of big tech and leaving regulators scratching their heads. Europe tried to rope them in, then thought better of it — realising regulating thousands of decentralised AI projects is like herding digital cats. The US applauds their ingenuity (while privately worrying about misuse), and China mostly ignores them, confident its firewall will filter out any troublesome innovation.
Meanwhile, governments scramble to tackle deepfakes, disinformation, and digital identity. Some are flirting dangerously with mandatory digital identity verification (“Know Your Customer” for the Internet), threatening to destroy online anonymity.
I speculated not long ago that a cracked or rogue ChatGPT in the hands of bad actors would lead to governments clamping down on the internet — more surveillance, speech restrictions, the works. That prospect is no longer hypothetical. Just look at how Western governments reacted to disinformation around elections and COVID — often overzealous, occasionally trampling civil liberties. Now multiply that fear by ten with AI in the mix (“Think of the terrorists using unhinged GPT!”) and you get the picture.
One particularly draconian idea gaining traction — as a “solution” to the AI trust problem — is “KYC for the internet.” In plain English it’s a “licence to surf”, forcing every internet user to verify their real identity (perhaps with government-issued ID) before posting or accessing certain content, to deter the armies of AI bots and deepfake trolls.
Australia’s political class — both the ruling Labor Party on the left and the Liberal-National Coalition on the right — have shamefully united to implement what’s effectively a “KYC for the Internet,” cynically justified as a measure to “protect the kids.” Particularly galling is that the Coalition, which has long claimed freedom of speech as a fundamental principle, enthusiastically backed this legislation. The ugly reality is that this law has little to do with child safety and everything to do with politicians wanting to silence Australians who use sock puppet accounts to say mean things about them online.
This concept, once confined to authoritarian playbooks, is now seeping into mainstream Western debate. Proponents argue it may be the only way to distinguish whether that Twitter account spouting nonsense is a human or an AI, the only effective method to hold someone accountable for maliciously deploying generative models. I get the temptation — the internet is already crawling with fake accounts, bots, and AI-generated disinformation, and it’s poised to become 100 times worse. But let’s call bullshit on this “solution” right now: sacrificing anonymity and privacy isn’t a real fix to AI misuse. Requiring every user to log in with a passport or fingerprint won’t solve the AI problem — it’ll just kill the free internet as collateral damage.
Yet, I fear that’s where we’re headed if nothing else works. Because frankly, the technical countermeasures are failing. We tried CAPTCHAs (“click the traffic lights to prove you’re not a bot”) — now AI can solve those captchas. We tried verifying users by having them take selfies holding ID — now AI can deepfake a video of that. We’re fast approaching a point where any digital authentication can be spoofed by sufficiently sophisticated AI.
As I grimly noted in an earlier talk, we may soon have no reliable way to authenticate humans online. That’s a societal crisis in the making. The response from governments will be to double down on verification: maybe some cryptographic signing of all content, maybe requiring AI-generated content to be watermarked (good luck with that). Or indeed forcing every user to attach a verified identity token to their actions. Each of those measures comes with enormous civil liberty implications and practical headaches.
All this regulation also has one fatal flaw: AI moves at warp speed, while bureaucracy crawls. By the time laws catch up, models have evolved beyond recognition. This ‘pace gap’ means we’re constantly regulating yesterday’s technology. The EU’s AI Act, for example, conceived when GPT-3 was state-of-the-art, won’t fully kick in until GPT-6 hits our screens. It’s like trying to regulate horses after the invention of the car.
So here we are: on the cusp of an internet where nothing can be trusted at face value — not text, not images, not even video of a person talking — and our leaders’ best idea is to effectively ask everyone to carry papers please to go online. Regulatory theater meets a technological pandemic.
I don’t have a perfect answer, to be fair. This is an unprecedented challenge. But I will say this: heavy-handed regulation driven by fear tends to overshoot and cause more damage than the problem itself. The chaotic flurry of AI rules globally is likely to produce a lot of unintended consequences and inconsistent outcomes, at least in the short term. Some jurisdictions will overshoot (stifling innovation), others will undershoot (fostering Wild West misuse). In between, businesses and users get whiplash trying to comply. The next two years will be incredibly messy on the regulatory front. Expect court challenges, trade disputes (“Your AI law gives your companies an advantage over ours!”), and companies occasionally just saying “screw it” and open-sourcing or moving offshore to avoid rules.
The next few years will be a rollercoaster as these competing regulatory philosophies collide head-on. AI nationalism is here, the rules are splintering, and innovation is racing ahead unchecked. Ultimately, the nations that strike the smartest balance — embracing innovation without destroying privacy and freedom — will shape the next decade. For now, it’s chaos, confusion, and endless compliance.
Welcome to the age of AI regulation.
The Synthetic Internet: A Collapse of Trust
In 2023, an AI-generated image depicting the iconic Hollywood sign engulfed by wildfires spread rapidly online. Completely fake, yet frighteningly believable, it fooled millions before being debunked. Welcome to the synthetic internet — where reality and fabrication blur seamlessly.
Beneath the dazzling marvels of AI, a deeper storm is brewing: our entire information ecosystem is collapsing. The online world is rapidly flooding with synthetic content — AI-generated articles, images, videos, and entire digital personas indistinguishable from authentic creations. In 2024, the trickle became a flood. By 2025, I’d wager a majority of the content online will be machine-made. I previously discussed the “dead internet theory” — the notion that much of the internet’s content is already fake or bot-generated — and we are now seeing it manifest in real time.
Take a simple example: social media. Scroll through your Facebook feed, and you might notice something is… off. Odd accounts posting weird images, like a slightly “uncanny” photo of the interior design of a room that looks almost real but not quite. Underneath, hundreds of generic comments from profiles with AI-generated profile pics.
Or the sudden surge of hyper-specific interest groups with active discussions that feel algorithmically on-the-nose. This isn’t a sci-fi story — it’s happening now. A large portion of social web activity is bot-driven, often for ad fraud or spam engagement. AI has supercharged this. Those bots now have endless plausible things to say, and endless fake images to share. They can fabricate entire communities out of thin air. It’s as if a digital hallucinogen has been dumped into our information water supply.
But the stakes extend far beyond spam and nuisance.
We’re entering a crisis of trust.
You come across what looks like a CNN news article — but it’s entirely AI-written and pumped out by a content farm for ad clicks. You see a video of a world leader making outrageous statements — but it’s a deepfake. You get a phone call from your daughter saying she’s been kidnapped and needs money — except it’s not your daughter, just an AI cloning her voice, which scammers used to prey on you.
All of these examples have already occurred in the last year. Now multiply them. The cost to generate misinformation or fraudulent content is trending toward zero, while the quality and realism of that content skyrockets. It’s the perfect storm for an epistemic crisis.
Check out some of the demos at Freelancer.com/ai and you’ll see how good an audio call with an AI agent can be.
We simply cannot trust our eyes and ears anymore when it comes to digital content. If all information becomes suspect, what happens? We risk a kind of collective schizophrenia, where public discourse deteriorates because no one knows what’s real. Already, scammers and propagandists are weaponising this uncertainty (“don’t believe anything unless we say it’s true”). We see authoritarian leaders dismiss authentic evidence as “deepfakes” to escape accountability, while conversely activists try to expose real deepfakes to prevent false panic.
The information war has escalated dramatically.
It’s like the boy who cried wolf, but with infinite wolves and infinite boys. Eventually the villagers just shrug and stop reacting to everything, real or not. Or worse, they believe whatever they want to believe, facts be damned, since objective truth is muddled. We’re already knee-deep in that era of subjective realities; AI-generated content could be the coup de grâce to consensus truth.
The tech companies are trying some measures. OpenAI built a (mostly ineffective) classifier to detect AI-written text — it basically didn’t work and they discontinued it.
Even Google couldn’t stop AI content being used for web content, initially telling everyone they would be penalised in the search ranking for using it before giving up — and if you can’t beat ’em join ’em — supplying AI content generation tools.
Researchers are working on watermarks for AI images and videos — but any such watermark can potentially be removed or spoofed, especially once multiple model architectures exist (a nefarious actor can just use a model that doesn’t apply the watermark). Efforts like the Content Authenticity Initiative (led by Adobe) aim to cryptographically sign legitimate content at creation (e.g., your camera signs a photo when you take it, proving its origin). Maybe that works for new content going forward, but it doesn’t solve the billions of legacy images and videos that can be repurposed, nor can it force bad actors to use hardware that signs their fakes.
I hate to say it, but for now the only rational stance is to assume everything digital could be a fake unless proven otherwise. That is a hell of a thing to live with. Imagine applying that standard to daily life — every email, every video call, every news item treated as guilty until proven innocent. It’s exhausting and corrosive.
In my MacroVoices discussions, I stressed how early signs of this were already visible. Blog spam and SEO content mills jumped on GPT-3 to crank out low-quality articles in 2022. By 2024, entire spam websites emerged that were AI-generated top to bottom, attracting clicks with sensational AI-written headlines. Even seasoned analysts were fooled on occasion by AI-generated research reports or fake interviews that looked real on first glance. The synthetic internet sneaks up on you; unless you’re actively looking, you may not notice that some of those Medium posts you read were written by GPT-4 with a bit of human curation.
The scariest part is when synthetic content is deployed deliberately to cause harm. We’ve seen state-aligned disinformation campaigns using AI to amplify their reach and tailor messages. It’s no longer a thousand trolls in a farm — it’s one troll with GPT-4 automating ten thousand fake personas.
Where it’s All Going — AI Predictions for 2025–2026
Model Commoditisation & The End of Proprietary Advantage
Open-source models are multiplying faster than rabbits, gutting the premium once commanded by closed systems like GPT-4. Soon, no serious business will pay high rents for generic model access — specialisation and application matter now. OpenAI will have a very challenging time. I suspect Altman will likely rage quit in frustration and start a new company while he still has the street cred, leaving what’s left of OpenAI to be devoured by Microsoft.
Big Tech Embeds AI Deeply, Like It or Not
Expect Microsoft, Google, and Amazon to ram AI into every piece of software you use — mandatory Copilot integration, AI-first Salesforce interfaces, the works. AI isn’t an add-on anymore; it’s your new default. Brace for embarrassing AI failures from companies bolting on AI without a clue which might lead to an emperor has no clothes moment and rush out of the cloud and into locally run AI at the edge- “not my AI not my data!”.
GPU Shortage Turns to Glut; Hardware Fragmentation Begins
Today’s Nvidia shortages become tomorrow’s oversupply as training demand plateaus and edge computing explodes. Meanwhile, China’s domestic silicon rises, smashing Nvidia’s near-monopoly.
Regulatory Chaos & the Dawn of “Verified Human” Internet
Brace for more messy AI legislation — likely from the EU first — that will slap big fines and send compliance departments scrambling. Watch for draconian attempts at digital IDs and content authentication and a huge backlash over it from the public. “Verified Human” badges are probably next, as your Instagram feed of unemployed models in bikinis gets replaced by legions of AI-generated thirst traps indistinguishable from the real thing.
AI Agents Transform Work — Humans Move Up the Stack
Forget static chatbots; AI agents become your new work colleague, slashing routine work like customer support and sales by double digits. Humans won’t vanish; instead, they’ll supervise AI or handle escalations. Say hello to roles like “AI workflow designer” or “AI auditor”. Personal assistants like Alexa finally grow a brain, doing tasks worth your time.
Netflix and the Infinite AI Content Glut
Netflix is about to discover exactly what Wikipedia felt like the day ChatGPT dropped — a once-great content king suddenly drowning in an endless ocean of AI-generated knock-offs, fan-fiction run wild, and synthetic spin-offs churned out at zero marginal cost. Good luck charging subscriptions when a thousand AI-generated spinoffs appear overnight, cooked up by teens with too much GPU and free time. We’ll finally get Game of Thrones Season 8 fixed, but also Season 9, 10, 11, in space, as a western, with you as the lead and all your friends as the characters.
Authenticity Crisis: The Internet’s Reality Check
As AI floods the web with convincing fakes, trust hits rock-bottom. Authenticity becomes a luxury — expect startups and platforms offering “verified real” content to explode. Local investigative journalism is already making a comeback because media organisations have trouble generating original content. Keep an eye out for a major deepfake-triggered crisis pushing governments to criminalise certain deceptive AI content, and using it as an excuse to gain more control over the electorate.
Economic Winners, Losers, and a Productivity Boost
Moderate productivity growth finally shows up in macro data, puzzling economists as inflation stays low despite wage hikes. Outsourcing to large-scale low-cost labour hubs shrinks as large employment locations like call centres become disrupted. Meanwhile, sectors like healthcare and education — previously tech-resistant — start to surge forward thanks to practical AI integration.
The AI-Powered Freelancer Explosion
AI has lit a fire under the global freelance market, elevating skills overnight and unleashing talent from every corner of the planet. Equipped with advanced tooling — think real-time AI copilots, multilingual translators, and expert-level writing and coding assistants — freelancers in emerging economies now deliver top-tier work at lightning speed, often indistinguishable from seasoned professionals in New York or London.
The old geographic premium for Western workers is quickly evaporating; businesses can now tap into a global talent pool that’s smarter, faster, and drastically cheaper. It’s the dawn of a borderless, hyper-competitive marketplace — adapt fast, or get ready to be outperformed.
All jobs move up the stack and an explosion of micro-entrepreneurship is both enabled and driven by the disruption caused by AI.
Societal Backlash and Political AI Battles
Just as organic food became chic, expect a cultural backlash: “Made by Humans” labels popping up everywhere. Politically, AI becomes a wedge issue — “AI stole your job” speeches incoming. Young folks adapt effortlessly; older generations push back, navigating thorny debates on what’s acceptable AI use. If GPT was allowed to write a prescription, half the work of GPs would disappear overnight and probably 70% of the work of therapists- with better outcomes.
The Black Box Problem: AI Keeps Surprising Us
Emergent capabilities continue baffling even the creators. AI’s black-box behaviour prompts intensified research into transparency and regulation. Brace for more “it taught itself to do what?” moments, fueling the push for tighter oversight. I definitely felt that moment again using Claude Sonnet 3.5 to write marketing copy and GPT 4.5 with Deep Research for, well, research.
Public Narrative Trails Reality — An Opportunity for the Smart
The average punter will underestimate how deeply AI is already embedded in daily life, giving savvy investors and technologists an edge. AI won’t disappear from headlines — it’ll just quietly become infrastructure, like electricity.
The AI storm rages on, reshaping everything from geopolitics to daily workflows. Time to plant your roots deep, adapt swiftly, and ride out what comes next.
Crazy Town
If there’s one thing I know for sure, it’s that the world is about to plunge headfirst into Crazy Town.
I mean, just wait until the Nigerian 419 scammers get full access to the latest GenAI tooling- real-time, high fidelity video-conference streaming with a GPT-4.5/Claude Sonnet 3.7 backend.
Nobody is going to pick up the phone anymore for unknown callers. You’ll need pre-shared, one time pad (passwords) to know if it’s truly them when they contact.
OnlyFans will become OnlyBots. Gaming won’t just be addictive; it’ll mutate into immersive, augmented realities where millions will happily trade their mundane daily existence for lives they could previously only dream of.
Who do you trust when mainstream media is pumping out state-sanctioned propaganda at scale, while the so-called “open” social media platforms are overrun by armies of AI-generated personas, drowning truth beneath synthetic noise?
At this point I wouldn’t be surprised if someone uses AI at scale to fake the second coming of Jesus Christ.
One thing I know for sure- Siri is still going to suck.
