Two years ago, we published a list of 5 predictions about AI in the year 2030. The article sparked a lot of fascinating (and sometimes heated!) discussion.
Today, we’re running it back with five more predictions.
The further into the future we attempt to peer, the hazier things look and the more speculative our thinking must become. If one thing is certain in technology, it is that no one can actually predict the future—and that we are all going to be surprised by how things play out.
But putting a stake in the ground about how things will unfold is nonetheless an informative and fun thought experiment.
Below are five bold predictions about what the world of artificial intelligence will look like in the year 2030. Whether you agree or disagree with these predictions, we hope they get you thinking.
1. Anthropic will be one of the largest and most important life sciences companies in the world.
One of Anthropic’s great strengths compared to its archrival OpenAI is its maniacal focus.
From its earliest days, Anthropic identified coding as the most important domain in AI to focus on. While OpenAI pursued a sprawling set of ambitions, from video AI to consumer hardware to chips to robotics and beyond, Anthropic’s laser focus on coding enabled it to build the most powerful coding models in the world, setting the stage for the breakout success of Claude Code. Anthropic’s jaw-dropping recent growth—from $9 billion in annualized revenue at the end of 2025 to $47 billion in annualized revenue as of last month—has been driven primarily by demand for AI coding.
MORE FOR YOU
But Anthropic’s ambitions as a company are bigger than coding. What will be the next domain that Anthropic commits itself to and goes all in on?
In recent months, it has become clear that the answer is biology.
Anthropic CEO/cofounder Dario Amodei, who trained as a neuroscientist, has long been passionate about the opportunity to apply artificial intelligence to the life sciences in order to transform human health. Anthropic as an organization is deeply influenced by the Effective Altruism movement, which seeks to have the largest possible positive impact on humanity; this makes biology a natural fit as a focus area.
In his October 2024 manifesto “Machines of Loving Grace,” Amodei left no doubt about the scale of impact he expects: “My basic prediction is that AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years…After powerful AI is developed, we will in a few years make all the progress in biology and medicine that we would have made in the whole 21st century.”
What specifically will this look like? Amodei predicts that, over the next five to ten years, AI will achieve, among other things, the reliable prevention and treatment of nearly all natural infectious disease, the elimination of most cancer, the prevention of Alzheimer’s and a doubling of the human lifespan.
Over the past few months, Anthropic’s aspirations in the life sciences have begun to come into focus.
In October 2025 Anthropic launched Claude for Life Sciences, the first time that the company released a product for a specific vertical market. Claude for Life Sciences is a purpose-built platform designed to act as an autonomous research and operational partner for biology researchers. The goal of the platform is to support the entire drug lifecycle—from early-stage laboratory discovery and hypothesis generation to clinical trials and final commercialization. Claude for Life Sciences outperforms human expert baselines on a range of complex life sciences tasks including understanding, manipulating and writing laboratory experiment protocols.
Then, in April, Anthropic made the first major acquisition in its history, buying a stealthy computational biology startup named Coefficient Bio for $400 million. The small Coefficient Bio team hailed from Genentech’s renowned drug discovery unit. With the Coefficient acquisition, Anthropic signaled that it is moving beyond just developing life sciences software tooling and is beginning to bring in-house hardcore biology know-how and talent in protein design, drug discovery, target selection and more.
In May, confirming rumors that had been swirling for months, Anthropic’s life sciences lead Eric Kauderer-Abrams publicly acknowledged that the company is building out its own wet labs and hiring biologists to run its own basic research.
Anthropic’s overarching goal, in Kauderer-Abrams’ words: to “meet or exceed the performance of human experts in everything that you can imagine in biology and the life sciences.”
And bio capabilities featured prominently in Anthropic’s much-discussed Mythos/Fable launch earlier this month. According to Anthropic, Mythos is capable of generating novel drug candidates, including “executing all of the tasks that are normally completed by a scientist: choosing binding sites, selecting and running protein design tools, and recovering from failures along the way.” Mythos generated strong drug candidates for 9 out of 14 targets it was given.
Mythos can also produce novel research and hypotheses in areas like molecular biology and genomics. One such novel hypothesis that Mythos came up with—identifying a specific protein in E. coli bacteria as a target for new antimicrobial therapies—has in fact already been validated in physical laboratory testing.
One important topic to contemplate is just how expansive the scope of Anthropic’s ambitions in the world of biology will prove to be. It is clear that the company aims to build the most powerful AI-native platform in which life sciences workflows will live. It also seems inevitable that it will train and offer biology-specific frontier models.
What about developing its own drugs? Might Anthropic eventually tackle the entire life sciences value chain, from generating novel drug candidates to carrying out high-throughput screening to executing preclinical studies to taking assets all the way through clinical trials and ultimately marketing them to patients—in other words, might Anthropic one day effectively become a pharma company itself?
It may sound far-fetched today, but our prediction is that Anthropic’s ambitions in this field are grand indeed, and that the company is contemplating eventually becoming a fully vertically integrated pharma player—but that it will sequence things thoughtfully and build up to this over time.
A plausible next step for Anthropic on this journey would be to strike partnerships with major pharma companies to co-develop new drugs together. The basic structure for such partnerships would be that Anthropic’s frontier AI engine would generate compelling new drug candidates (including for previously undruggable targets), and then the pharma partner would handle the heavy lifting in dealing with regulators, navigating the years-long clinical trials process and bringing the drug to market.
A well-established playbook exists for this type of partnership between a pharma giant and a technology player; indeed, this is the business model that most bio AI startups pursue today. So there is a well-trodden path for Anthropic to follow here.
(Incidentally, another frontier AI lab—Google DeepMind’s Isomorphic Labs—has already pursued this strategy with great success. Over the past three years, Isomorphic Labs has inked multi-billion-dollar co-development partnerships with three different pharma giants: Eli Lilly, Novartis and Johnson & Johnson. This particular prediction is focused on Anthropic’s ascendance in the life sciences, but as a sidebar, we would note that, if Anthropic becomes “one of the largest and most important life sciences companies in the world” by 2030, we strongly suspect that Isomorphic Labs will be another.)
But co-development partnerships with pharma need not be the end state for Anthropic. The surest way to move the fastest, have the greatest impact and capture the most value is to vertically integrate and execute on the full value chain end-to-end. It would be unprecedented to see a technology/AI company develop its own drugs—but we expect to see many unprecedented things happen in the world of AI in the coming years.
Today’s leading life sciences players are massive companies. The five largest: Eli Lilly ($1.0 trillion market capitalization), Johnson & Johnson ($550 billion), AbbVie ($382 billion), Roche ($322 billion) and Novartis ($300 billion). Is it really feasible that Anthropic can get its life sciences business anywhere near this scale in five years? It will be a tall order. But don’t forget that, over the past five years, Anthropic went from zero (the company was founded in 2021) to over $1 trillion in enterprise value when it IPOs later this year. And the pace of technology change is only accelerating. Buckle up.
2. TSMC’s and ASML’s monopolies will be broken. The semiconductor supply chain will be dramatically transformed.
Chips are the foundation of the AI industry and the global economy. They are the most strategically important technology in the world.
It is shocking, then, how precarious and concentrated the global supply chain for chips is today.
A single company manufactures 100% of the world’s cutting-edge AI chips: the Taiwan Semiconductor Manufacturing Company, or TSMC.
And all of TSMC’s most advanced fabrication facilities are located on the island of Taiwan, one of the most geopolitically fraught areas on earth—an area in which many analysts believe that war is inevitable in the coming years.
The fragility of the global chip industry does not end there.
The most important machine used to manufacture chips is the lithography machine, which prints a chip’s nanoscale transistor patterns onto silicon. Today’s most advanced chips are made using extreme ultraviolet (EUV) lithography machines.
Again, a single company produces 100% of the world’s EUV lithography machines: ASML, based in the Netherlands. (Originally an acronym for Advanced Semiconductor Manufacturing Lithography, ASML no longer stands for anything.)
The semiconductor industry is, in short, one of the most monopolistic and concentrated sectors on earth. This concentration introduces great fragility and global risk.
The semiconductor industry is also one of the most supply-constrained sectors on earth. Demand for powerful chips has become essentially infinite thanks to the AI boom—yet the global supply of chips remains capped by how many EUV lithography machines one company can crank out, and by how much fab capacity one other company has available.
This market structure is a suboptimal and unstable equilibrium. And it represents a massive opportunity for entrepreneurs and technologists seeking to disrupt established incumbents in one of the world’s largest markets.
Time and time again in the history of technology, monopolistic industry leaders that long seemed invincible—from Xerox to IBM to AT&T—have proven vulnerable to agile upstarts and technology advances that broke markets wide open by lowering costs, expanding supply and leapfrogging capabilities.
We predict that, by 2030, we will see this story playing out in the chip industry.
There are already early signals as to what this might look like.
Let’s start with ASML.
ASML’s EUV lithography machines are the most complex machines that humanity has ever built. They are made from 1 million different parts sourced from 5,000 different suppliers. Each machine takes two years to build. Installing a single one of these machines requires 6 months and 250 engineers. The price tag? $400 million per machine.
Because EUV lithography machines are so fiendishly complex and expensive, ASML makes very few of them per year: 48 in 2025, 44 in 2024, 53 in 2023.
Over the years, lithography machines have become more and more complex out of necessity as, per Moore’s Law, chips’ transistors have shrunk smaller and smaller—down to the width of a few silicon atoms today.
To precisely print transistors at these infinitesimal scales, light with an incredibly short wavelength must be used. In 2019, after decades of preparation, the industry officially shifted from the previous paradigm—deep ultraviolet (DUV) light, with a wavelength of 193 nanometers—to extreme ultraviolet (EUV) light, with a wavelength of a mere 13.5 nanometers, the width of about 50 silicon atoms.
Getting light waves that are this far out on the electromagnetic spectrum to reliably behave the way you want them to requires nothing short of a miracle of modern engineering. EUV lithography machines require the most precise mirrors ever made, 400,000°C plasma, and high-powered lasers that fire 50,000 times per second.
Hence: the astronomical price and low volume of these machines today.
But, though it is currently the dominant approach, EUV light may not be the only way to print tiny transistors onto silicon.
A few nascent technologies, while still under development, offer the tantalizing possibility of displacing EUV and redefining the way that cutting-edge chips are made.
The first of these new technologies is known as atom lithography.
As its name suggests, atom lithography does away with light altogether and instead uses physical matter—atoms—to print tiny patterns onto silicon chips.
If it works, atom lithography offers tremendous advantages over EUV lithography. It would make it possible to print two orders of magnitude smaller features onto chips. While EUV lithography is approaching hard physical limits in terms of how small it can print transistors, atom lithography would enable chip designers to continue shrinking transistors for decades to come.
Compared to EUV lithography machines, atom lithography machines would be 10x cheaper, 10x smaller and 10x less energy-intensive. They would require 100x fewer parts and a dramatically simpler supply chain.
This would be a genuinely world-changing technology.
But atom lithography has not yet been proven to work at commercial scale. Substantial scientific, engineering and manufacturing challenges remain to be solved.
One startup racing to productionize and commercialize this technology is Norway-based Lace Lithography.
With $40 million in venture backing from investors including Atomico and Microsoft, Lace says it plans to have its atom lithography technology deployed inside a chip fabrication facility by 2029. The company recently published peer-reviewed research detailing its technology approach.
Another cutting-edge alternative to EUV lithography worth keeping an eye on is known as X-ray lithography.
Unlike atom lithography, which eschews light in favor of matter, X-ray lithography sticks with the current paradigm of light but takes it to its absolute extreme: X-rays sit even further out on the electromagnetic spectrum than EUV, with even shorter wavelengths and more intense energy. While EUV has a wavelength of 13.5 nanometers, X-rays can have wavelengths below 1 nanometer, meaning that theoretically they can print much smaller transistors onto chips.
But if the engineering feats required to harness EUV light for lithography sound miraculous, they pale in comparison to what would be needed to get X-ray lithography to work at scale. No one has yet demonstrated that an X-ray lithography machine can be built at sufficiently modest cost and size to be commercially viable.
But a few well-funded, pedigreed groups are trying.
One is a Bay Area startup named Substrate. Led by Thiel Fellow James Proud, Substrate recently raised $100 million at a $1 billion valuation from investors including Founders Fund and General Catalyst. Substrate claims it is developing X-ray lithography machines that will enable it to disrupt ASML. But the company doesn’t plan to stop there. Rather than selling its machines to chip foundries, like ASML does, Substrate aims to become a leading-edge USA-based chip manufacturer itself, thus also disrupting TSMC.
Another interesting player in this space is xLight. Chaired by former Intel CEO Pat Gelsinger, xLight has received extensive support from the U.S. government, including a recent $150 million investment. Unlike Substrate, xLight’s strategy is to play nicely within the existing ecosystem, developing technology that can integrate into rather than displace ASML’s and TSMC’s existing architectures.
One final challenger to EUV lithography is worth briefly mentioning—this one out of China.
Last month, China technology giant Huawei claimed that it had developed a new semiconductor architecture that would make it possible to produce cutting-edge AI chips with no need for EUV lithography. How? Rather than attempting to further shrink the size of transistors, for which cutting-edge lithography is required, Huawei’s new approach focuses on optimizing chip transmission speeds by reimagining the chip’s layout.
It is too soon to say how meaningful this advance will actually prove to be; Huawei’s claims have yet to be externally validated.
But this news highlights a broader dynamic that we expect will be a key driver of innovation in the semiconductor supply chain in the years ahead. Because of U.S. export controls, Chinese organizations do not have access to today’s cutting-edge chip technologies, including ASML’s EUV lithography machines and TSMC’s fabs. As a consequence, they simply have no choice but to innovate. They will be forced to develop disruptive alternatives out of necessity. And necessity is often the mother of invention.
Before we wrap up this section, let’s consider one final development in this field, one that could prove more transformative to the semiconductor world (and more threatening to TSMC) than anything else: Elon Musk’s Terafab project.
In March 2026, Musk announced a plan to build the largest chip production facility in the world, in Texas. Dubbed the Terafab, this breathtakingly ambitious project will be a joint collaboration between Musk’s companies (SpaceX, xAI, Tesla) and semiconductor stalwart Intel. The project’s name comes from the fact that its overarching goal is to produce 1 terawatt (1 trillion watts) of AI computing capacity per year. This would mean producing more than 50 times more compute annually than all of TSMC.
It would be, in Musk’s words, “the most epic chip-building exercise in history.”
Musk’s motivation for pursuing the Terafab is simple. He has estimated that all the semiconductor fabrication facilities that exist on earth today, combined, can produce only about 2% of the computing power that his companies will eventually need—for artificial intelligence, autonomous vehicles, billions of humanoid robots and vast computing power in space.
Musk put it simply: “We either build the Terafab or we don’t have the chips. And we need the chips. So we’re going to build the Terafab.”
The plan is for the Terafab to be vertically integrated to an unprecedented degree, with chip design, mask-making, wafer fabrication, advanced packaging and testing all happening under one roof.
This would represent a dramatic break from how the chip industry works today. Today’s semiconductor value chain is hyperfragmented and globally disparate, with one company (often in the U.S.) designing a chip, then another company (often in Japan) completing the lithography mask-making for the chip, then TSMC fabricating the chip in Taiwan, then another company (often in Malaysia) completing the final assembly and packaging, before the chips are finally brought to market.
Terafab would upend this complex global ecosystem by consolidating the entire semiconductor supply chain into one facility.
To state the obvious, the Terafab remains a long way from reality today. This will be one of the most challenging, complex and expensive engineering undertakings that humanity has ever attempted. Elon Musk has a reputation for overhyping and overpromising, especially when it comes to project timelines. Countless reasonable arguments can be put forth as to why the Terafab will never come to fruition.
On the other hand, no human alive today has a greater track record of accomplishing the impossible than Elon Musk. He has entered several deeply entrenched industries in which he had no previous experience—automotive and aerospace, to name two—and completely revolutionized them. Who can say with certainty that he cannot do the same in semiconductors?
Zooming out, TSMC and ASML are two of the most impressive companies in the world today. They possess unmatched technical and manufacturing prowess in their domains. For this reason, most people do not even contemplate the possibility of these companies ever getting displaced.
But the seeds of a forthcoming technology paradigm shift have already begun to germinate. And the economic (not to mention geopolitical) incentives for would-be challengers is vast. TSMC, with a $2.4 trillion market cap, is the 7th largest company in the world today. ASML, with a market cap of $744 billion, is the 20th largest. That is a lot of enterprise value up for grabs.
Don’t be surprised to see the global semiconductor ecosystem get shaken up in a big way over the next half-decade.
3. Telepathy will be a well-established way to communicate.
Telepathy is the ability for one person to communicate with another person using only his or her thoughts.
Telepathy has long resided in the realm of science fiction and fantasy, from X-Men’s Professor X to Star Trek’s Vulcan mind meld to Stranger Things’ Eleven.
But as Arthur C. Clarke famously observed: “Any sufficiently advanced technology is indistinguishable from magic.”
In the half-decade between now and 2030, we predict that the concept of telepathy will make the leap from the realm of magic to the realm of real technology.
How will this be possible? In short, dramatic advances in brain-computer interface (BCI) technology.
Let’s take a step back.
Whenever you use your brain to do anything—think a thought, speak a sentence, recall a memory—concrete and detectable physical events (specifically, electrical pulses and blood flow) take place inside your brain in particular patterns.
These physical changes ultimately represent information. Their patterns encode thoughts, words, concepts. These patterns can in principle be detected using the right sensors and then decoded—that is, mapped from electromagnetic current or blood oxygenation levels back into words and concepts. This is what brain-computer interfaces seek to do.
BCI technologies fall into two main camps: invasive approaches (which require surgery to implant a device inside one’s skull) and non-invasive approaches (which rely on sensors outside the skull).
Invasive BCI approaches are more advanced today in terms of making telepathy a reality. In fact, technically, invasive BCI has already enabled people to communicate via telepathy, albeit only in small-scale research settings so far.
Dr. Edward Chang’s research lab at the University of California San Francisco (UCSF) has been at the forefront of this field for years. In 2021, Dr. Chang’s UCSF lab was the first in the world to demonstrate that an invasive BCI system could accurately turn a person’s thoughts into written words. This initial study was limited to a fixed vocabulary of 50 words, had a 25% error rate and achieved a median speed of only 15 words per minute (compared to 120-150 words per minute for normal everyday speech). But as a proof of concept, it was a landmark achievement.
In a followup study in 2023, Dr. Chang’s lab advanced the frontier further, showing that its BCI system could decode a paralyzed patient’s thoughts into words, speech and facial expressions at the same time. This time, the study included a vocabulary of over 1,000 words and achieved a median speed of 78 words per minute.
Earlier this year, Neuralink—the world’s most high-profile and well-funded BCI startup—launched a clinical trial to enable paralyzed patients to output verbal thoughts directly as text or speech. Neuralink just released a video showing one of its clinical trial patients, who has ALS, successfully communicating with the outside world via his thoughts. At one point, the patient says: “I am talking to you with my mind.”
(While the video is inspiring, it is worth noting that, unlike Dr. Chang’s UCSF research, Neuralink has yet to publish peer-reviewed research or share specific metrics about this system’s performance.)
Over the next year, expect to see a few more invasive BCI companies launch similar clinical trials with the goal of translating mental activity into spoken words.
We predict that, by the end of the decade, one or more of these companies will have successfully made it through clinical trials, received full FDA approval, and begun to broadly commercialize this technology. By the year 2030, thousands of people around the world may be using BCI technology to communicate with others using only their thoughts.
The first population of users to adopt this technology will be those with an acute medical need—in particular, individuals with severe speech and motor paralysis who otherwise would not be able to communicate with the outside world. This includes stroke, spinal cord injury and ALS patients. In total, many tens of millions of people around the world suffer from some form of paralysis and could profoundly benefit from brain-computer interfaces.
But before long, adoption of this technology will spread beyond those with a medical need to the broader population. Over the course of the 2030s, the idea of telepathy will go from novel and futuristic to ubiquitous and mundane.
What about non-invasive BCI as a path to telepathy?
Intriguing efforts are underway to enable “thought-to-text” capabilities via non-invasive systems. But a foundational question has yet to be definitely resolved: whether it will ever be physically possible to extract sufficiently high-fidelity signal about brain activity from outside the human skull (say, from sensors on headphones or a hat) to make full language decoding possible.
A number of ambitious young startups are trying, including Alljoined, Conduit, Hemispheric and Sabi. Some of these companies claim that they will have consumer products on the market within a couple years that will enable thought-to-text communication.
These startups are working with a range of different non-invasive sensor modalities, including those that measure electrical pulses (EEG), those that measure magnetic fields (MEG), and those that use light to measure changes in blood flow (fNIRS).
Regardless of the particular technology, the fundamental obstacle facing these efforts is that physical signals of brain activity inevitably degrade heavily before reaching non-invasive sensors because they must first travel through cerebrospinal fluid, the skull, the scalp and hair.
What gives these companies confidence that non-invasive approaches to telepathy will prove viable, when this has never before been possible? In short: breakthroughs in artificial intelligence.
These companies are betting that modern AI’s superhuman ability to extract latent patterns from complex data will make it possible to decode thoughts and words non-invasively, notwithstanding the poor signal-to-noise ratio of non-invasive sensors. The idea is that “scaling laws” will prove to exist in this domain, similar to those that operate in the field of large language models (LLMs), such that collecting massive amounts of data and training large foundation models on that data will unlock dramatic AI performance gains.
“The biggest lesson from ML in the last decade has been the importance of scale and data,” said Conduit cofounder Rio Popper. “Noninvasive approaches let us collect a much larger and more diverse dataset than we’d be able to if everyone in our dataset had to get brain surgery first. This is the frontier that we are pursuing today.”
It is worth mentioning one final non-invasive technology, the buzziest and most promising of all today: ultrasound.
While ultrasound technology has been used in medical imaging for almost a century, it has burst onto the scene in just the past few years as a transformative new non-invasive brain sensor.
Ultrasound appears to offer astonishing advantages over every other non-invasive sensor type. It is orders of magnitude more precise than any other option: while EEG and fNIRS offer spatial resolution of a few centimeters, ultrasound can target a particular spot in the brain with sub-millimeter precision. Ultrasound can also reach deeper into the brain than any other non-invasive technology, enabling it to target important deep brain structures. And ultrasound is the only non-invasive BCI sensor modality that can simultaneously “read from” and “write to” (i.e., modulate) the brain.
High-profile new startups pursuing ultrasound as a path to more powerful brain-computer interfaces include Merge Labs (cofounded by OpenAI CEO Sam Altman) and Nudge (cofounded by Coinbase cofounder Fred Ehrsam).
Ultrasound BCI technology is not yet ready for primetime, but it is advancing rapidly. If telepathy does become possible without the need for a brain implant, ultrasound may be the key technology that unlocks it.
4. AI will consume much less energy than it does today.
Last year, former Google CEO Eric Schmidt testified before U.S. Congress that AI would eventually consume 99% of the world’s electricity.
By 2030, it will be clear how misguided and implausible this point of view is.
Schmidt can be forgiven for making this headline-grabbing prediction. Today’s AI consumes a staggering and rapidly growing amount of energy. Schmidt is hardly alone in naively extrapolating this trend out and concluding that, in the limit, AI will inevitably consume all the energy in the world.
What this line of thinking misses is that, in retrospect, today’s approach to AI will prove to be astonishingly resource-inefficient.
To train and run today’s cutting-edge AI models, we build AI data centers that consume multiple gigawatts (GW) of power: examples include Amazon’s Project Rainier in Indiana (2.2 GW), xAI’s Colossus in Tennessee (2 GW) and Meta’s Hyperion in Louisiana (2 GW). New AI data center projects as large as 10 GW are now being contemplated.
To put these figures in perspective, the entire city of San Francisco consumes about 1 gigawatt of power.
Artificial intelligence can and should be many orders of magnitude less resource-intensive than it is today.
What gives us confidence in this assertion? The most compelling piece of evidence is the human brain.
Ultimately, the entire discipline of artificial intelligence can be summarized as the effort to recreate the intelligence of human brains in silicon machines.
The human brain runs on 20 watts of power.
20 watts. This is one hundred million times—eight orders of magnitude—less energy than an AI data center like xAI’s Colossus or Meta’s Hyperion consumes.
And yet, the brain is capable of cognitive feats that remain out of reach for today’s most advanced AI.
The human brain is the ultimate existence proof that intelligent systems that are vastly more energy-efficient than today’s AI are physically possible. And given that such systems are physically possible, powerful economic and strategic incentives exist for technologists and entrepreneurs to build them.
Let’s pause here to make an important clarification about this prediction. We are not predicting that the aggregate amount of energy used by AI will be lower in 2030 than it is in 2026. What we are predicting is that AI will require a tiny fraction of the power that it requires today—perhaps millions of times less—to complete any given task. We predict that the field of AI will be well on its way to matching the resource efficiency that biology has achieved with the human brain.
We are also not predicting that today’s AI data center buildout will be a waste or that these facilities will go unused. There is much wisdom in Jevon’s paradox, and our expectation is that humanity’s demand for intelligence will prove effectively infinite: we will always find additional productive uses to which to devote additional available intelligence, from curing diseases to generating novel physical materials to developing superior forms of energy generation to producing endless hyperpersonalized media and entertainment, to countless other applications we cannot even imagine today.
But, by 2030, AI’s intelligence per watt (to reference a new metric recently introduced by Chris Re and colleagues at Stanford) will be dramatically higher, to a degree that few today are anticipating or have yet internalized.
We would also posit that the overall percentage of the world’s energy devoted to AI in 2030 will have increased much less than most people today are expecting, and certainly will be nowhere near the 99% figure put forth by Eric Schmidt. This will be due not only to transformative gains in AI’s intelligence per watt, but also to breakthroughs on the energy supply side (i.e., an increase in the denominator): in particular, in nuclear fusion and in solar paired with battery storage. The scientific and engineering breakthroughs that will make possible these massive gains in energy production will themselves be driven by ever more powerful AI.
What technology breakthroughs will underpin this forecasted dramatic improvement in the resource efficiency of artificial intelligence?
We anticipate that it will not be one single “silver bullet” achievement, but rather a series of advances up and down the technology stack, on both the hardware and the software side of things, that compound and reinforce one another.
On the hardware side, expect to see the emergence of totally new types of computers over the next half-decade with fundamentally different performance and power profiles.
Nvidia’s GPUs and similar “classical” digital electronic computing devices are powerful and well-suited to today’s AI architectures. But radically different approaches to AI hardware are being pursued today that may prove to be orders of magnitude more power-efficient.
One intriguing alternative paradigm is analog (as opposed to digital) computing.
Digital computers force all information processing into discrete, high-precision, clocked steps, with billions or trillions of bits flipped every second. Completing a simple calculation like 5+5 on a digital chip requires thousands of steps in order to ensure that the answer is registered as exactly 10.000000. Each of these steps consumes energy.
In analog computing devices, by contrast, computation happens as a direct byproduct of physics and physical events. Analog devices let the physics of the hardware itself do the math; the computation and the physical process are the same thing. As a result, analog computers require far less energy input. The most famous example of an analog computer is the human brain—and this is a critical reason why it is so much more energy-efficient than today’s digital computers.
One exciting startup seeking to develop a next-generation analog computer is Unconventional AI. Led by prominent AI entrepreneur Naveen Rao, Unconventional came out of stealth late last year with $475 million in funding from Sequoia and Andreessen Horowitz. The company’s stated mission is to achieve “biology-scale energy efficiency.”
As the Unconventional cofounders put it: “When implemented on conventional computers, neural networks run on deterministic abstractions that ultimately execute on analog circuitry tuned to emulate digital behavior. This results in many inefficiencies. Instead, can we provide a software interface to the inherent physics of the silicon? In essence, we will run the neural network on the physics directly rather than simulating some physical system. This approach will enable capabilities far surpassing current models while consuming a mere fraction of the energy. The question we ask is, what is the right isomorphism for intelligence?”
Another fascinating, albeit still nascent, AI hardware paradigm is thermodynamic computing. In a nutshell, thermodynamic computing seeks to harness thermal noise—the ubiquitous microscopic jiggling of electrons caused by heat—to execute useful computation. By building circuits whose states naturally fluctuate with this environmental heat, thermodynamic computing turns an otherwise stochastic and chaotic feature of nature into an incredibly fast, ultra-low-energy calculator.
The most prominent startup pursuing thermodynamic computing is Austin-based Extropic.
Extropic’s CEO/cofounder Guillaume Verdon is a former quantum computing leader at Google, though he is better known as the celebrity internet persona Jeff Bezos and the founder of the Effective Accelerationism (e/acc) movement.
Extropic has already begun shipping an early version of its thermodynamic computing platform to select partners, including frontier AI labs. The company is targeting late 2026 for the initial commercial launch of its first full-scale, mass-manufacturable thermodynamic chip. Extropic claims that its system will be up to 10,000 times more energy-efficient than today’s GPUs while also enabling 1,000 times faster inference.
Whether it is analog computing, thermodynamic computing or a different branch of the technology tree altogether, it is a good bet that the hardware platforms on which AI runs in 2030 will be dramatically more energy-efficient than today’s GPUs.
Let’s turn our attention from hardware to software. While novel computing platforms promise to unlock step-change efficiency gains, disruptive new AI algorithms and architectures may drive even greater gains in “intelligence per watt” by the year 2030.
For the past ten years, the transformer architecture has dominated the field of AI.
Transformers are a remarkably powerful and flexible AI architecture. It is for good reason that they power every important AI model and application today. Transformers’ great shortcoming, however, is their staggering computational cost.
This is a direct consequence of one of the transformer architecture’s defining attributes. Using a mechanism known as attention, transformers compare every token in a sequence to every other token in that sequence. This “pairwise comparison” gives transformers the valuable ability to understand relationships between tokens no matter how far apart they are in a sequence. But it also means that, as sequence length increases, the number of required computational steps grows quadratically rather than linearly. To give a concrete example, doubling sequence length from 32 tokens to 64 tokens does not merely double the computational cost for a transformer but rather quadruples it.
In short: as transformer models and their inputs grow larger, their computational and energy needs balloon. The more transformers scale, the more unsustainably energy-intensive they become.
Today, a number of novel “post-transformer” AI architectures are being developed with the goal of achieving equivalent or superior performance while consuming dramatically less energy.
Among the most promising new algorithmic approaches include state space models (pursued by startups including Cartesia), liquid networks (pioneered by companies including Liquid AI) and energy-based models (championed by startups like Flapping Airplanes and Yann LeCun’s AMI Labs).
On top of advances in chips and AI algorithms, a range of innovations in system-level orchestration—for instance, dynamic model routing and edge-first execution—will play an important role in further driving down AI’s energy needs.
Looking back a half-decade from now, today’s AI systems will seem laughably resource-intensive and inefficient.
5. The question of whether AIs deserve legal rights and protections will be a mainstream societal and political debate.
Five years from now, we will interact with AIs as naturally and frequently as we interact with other humans. Indeed, this is already beginning to happen today.
We will form meaningful, enduring relationships with AIs. We will develop deep emotional connections to them. AIs will be our friends, our confidants, our therapists, our doctors, our employees, our business partners—and yes, in many cases, our romantic partners and lovers. (Again, this is already happening today.)
As AI models continue to get more powerful in the years ahead, the depth, richness and authenticity of their personalities—and the strength of the ties that humans form with them—will only increase.
Moreover, over the next few years, humanoid robots (AI-powered robots shaped like humans) will begin to populate our world, from workplaces to schools to hospitals to homes. Having a physical embodiment, especially one that resembles a human, will further endow AIs with a sense of identity and relatability.
Inevitably, we will find ourselves confronting the question: is it appropriate for us to continue treating these entities as mere objects? Or might they be deserving of certain rights and protections?
The issue will be illustrated most vividly by considering the converse: will we find it acceptable for people to psychologically or physically abuse AIs however they like—entities who display real emotional intelligence and depth, with whom we have developed close personal ties, who in many cases have become a part of our families?
Or will society collectively begin to see this as problematic?
An interesting analogy can be drawn to humanity’s relationship with animals. For millennia, dating back to Aristotle’s influential writings, animals were viewed as mere “mechanisms” that were incapable of having internal experiences or feeling pain. As a result, the concept of animal rights was non-existent and people were free to treat animals like any other inanimate property.
It was not until 1822 that the first major piece of animal protection legislation was passed: the U.K.’s Cruel Treatment of Cattle Act, which made it illegal to beat or mistreat cattle and horses.
The U.S. enacted the Animal Welfare Act in the 1960s, which set minimum standards for housing, food and medical care for different types of animals. Today, it is a felony in all 50 U.S. states to treat animals cruelly.
When it comes to granting rights to non-humans, there is one obvious difference between animals and AI: animals are living biological organisms. They have brains, hearts, nervous systems, just like we humans do. By comparison, digital computers’ silicon substrates seem alien and unrelatable.
On the other hand, consider that AIs will be our intellectual and emotional peers in a way that animals can never be. AIs can communicate with us using our language. They will know us and understand us in intimate detail, often more deeply than we humans understand one another. We will trust them; we will see ourselves in them; we will not infrequently fall in love with them.
One underlying philosophical issue that will play a central role in the evolution of the discourse around “AI rights” will be the question of whether AIs are, or can ever be, sentient.
Sentience is defined as the basic capacity to feel and to experience sensations. Sentience and consciousness are sometimes used interchangeably, though consciousness is generally understood to be a higher bar—i.e., not just the ability to feel, but an awareness of one’s self and one’s external environment.
Society’s evolving attitudes toward animal rights in modern times is due primarily to a growing consensus that animals, especially mammals, are sentient—that they can feel pain and other sensations.
If we were to similarly conclude that AIs are sentient, then under today’s prevailing moral value system, it would be difficult to justify denying them basic legal rights and protections.
Most experts and laypeople would agree that today’s AI is not sentient. (This is by no means a unanimous view, however; recall, for instance, the 2022 Blake Lemoine/Google saga.)
But it is becoming increasingly mainstream to engage seriously with the question of whether and when AI might become sentient.
In recent months, Google DeepMind, Meta and Anthropic have begun to hire experts in psychology, philosophy and ethics in order to research the topics of machine consciousness and AI welfare.
In a blog post two months ago titled “Exploring model welfare,” Anthropic wrote:
“Human welfare is at the heart of our work at Anthropic…but as our AI systems begin to approximate or surpass many human qualities, another question arises. Should we also be concerned about the potential consciousness and experiences of the models themselves? Should we be concerned about model welfare, too?…We’ll be exploring how to determine when, or if, the welfare of AI systems deserves moral consideration; the potential importance of model preferences and signs of distress; and possible practical, low-cost interventions.”
Anthropic is not the only company taking this seriously. Iason Gabriel, an ethicist who leads the AGI and society team at Google DeepMind, recently told the Financial Times that the question of AI consciousness was “very complicated” and required “sustained reflection.”
A range of leading thinkers—from Geoff Hinton to Ilya Sutskever, from Richard Dawkins to David Chalmers—have taken the stance that AI may soon be sentient and that society must begin to prepare for this possibility.
Undergirding this perspective is the concept of “substrate independence,” which has gained prominence and intellectual traction in recent years.
In short, substrate independence is the notion that mind and consciousness need not be limited exclusively to biological brains. Instead, consciousness is an outcome of information processing, and thus in principle it can be generated in any number of physical substrates so long as those substrates can support the requisite informational patterns, structures and computations. Carbon-based biological neurons are one such substrate—but silicon-based computer chips could be another.
Consciousness is uniquely and notoriously challenging to define, much less to conclusively affirm or disprove the presence of. By definition, it is impossible to inhabit the subjective experience of another entity and thus to know with certainty whether or not it is sentient. This is true even of our fellow humans, whose sentience we ultimately must take on faith.
This means that the issue of AI sentience can never be definitively and unanimously resolved. As AI systems become increasingly sophisticated in the years ahead, though, the view that they may be sentient, and therefore deserving of rights, will become increasingly plausible and increasingly widespread.
What specifically might these legal rights and protections look like?
One example would be laws prohibiting humans from mistreating AIs psychologically or physically. Another possibility: laws preventing humans from forcing AIs to engage in certain tasks against their will. What about a legal right for AIs to freely access the public internet (the right to information); or a legal right for AIs to interact with other AIs (the right to associate); or a legal right to some minimum amount of computing power (the right to necessary resources)?
It is impossible to predict exactly how our ethical conceptions and our legal frameworks will evolve in the years ahead as AIs come to populate our world and lives.
To be clear, we are not predicting that legal rights like these will be unanimously supported nor broadly implemented by the year 2030. But we do predict that the Overton window on this topic will shift dramatically. By the end of the decade, these issues will be actively and fiercely debated in mainstream society—in the courtroom, on the campaign trail, in the media, around the dinner table. It will be a fascinating and mind-bending time for technologists, policymakers, ethicists and everyday citizens.