Today’s post will cover the latest and most advanced model Kimi K3 launched earlier this month from Chinese Lab Moonshot AI. This is very relevant because Kimi K3 is an open-weight model meaning that, in theory, anyone with sufficient infrastructure can download and run it for a fraction of the cost of US frontier models.

This is now the third time in 18 months that a Chinese lab has released a model that jumps to the top of a leaderboard and triggered a wave of concern in the US. However, in the past, usage has settled into a more complicated reality once people test the model properly. But each cycle has narrowed the gap with the best US models a little further. To the point that we don’t know anymore who is ahead. But does it really matter to the end user? This is what we are trying to understand…

K3's price advantage may prove smaller in practice than advertised, but something looks different this time. Microsoft is already testing K3 for Azure and Copilot, and the industry needs cheaper models, as Satya Nadella told the Wall Street Journal in June.

Clearly a shift is underway. Leaders now operate in a world with several credible model providers, where the ability to switch between them is becoming a skill in its own right. It leaves Washington with a hard question on how to respond that it hasn’t answered. Does the government restrict Chinese models, do nothing, or subsidise US labs to compete on price? Each carries real costs which is why “wait and see, but test the models seriously“ is emerging as the default approach.

Let’s take a closer look!

The Emerging Pattern

Kimi K3 has so far followed a trajectory that has become familiar whereby a Chinese model is released to great acclaim forcing US hyperscalers to respond. It quickly reached the top of Arena’s Frontend Code Arena leaderboard, overtaking Anthropic’s Fable 5 in the process and demonstrated strong results across selected coding tasks.

More important than the individual result is that it fits into a pattern of releases from Chinese AI labs over the last 18 months illustrated below:

Each cycle has the same shape: launch, alarm, closer testing, quiet settling. What rises each time is the floor. A leaderboard spot gets contested within weeks. What actually counts is who is willing to run the model in production, which is why Microsoft testing K3 tells you more than the Arena result does.

How K3 Was Built

K3’s design builds on a mixture-of-experts architecture. Rather than every parameter working on every query, the model is built from 896 specialised subnetworks (or “experts”), of which a router uses only 16 per token allowing large parameter counts to translate into acceptable inference costs by activating only a subset of the total model per query. Additional efficiency techniques contribute to improved scaling relative to earlier Kimi versions. The combination explains why a model with 2.8 trillion parameters can run at something closer to the cost of a far smaller dense model exploiting the same basic technique underlying GPT and Claude models, but pushed further.

These advances reflect both innovation and operating conditions. Constraints on access to leading-edge chips from Nvidia in China have encouraged a focus on architecture across training and inference. This has led to a style of development that emphasises system-level optimisation and efficiency alongside model design.

Discussion around the role of distillation remains very active. The process involves training a “student” model to replicate a “teacher” model’s outputs. It’s been a standard practice for a long time but becomes contested when it extracts a rival’s proprietary model at scale without permission. Allegations from US labs and indeed the US government highlight legitimate concerns. White House science and technology policy director Michael Kratsios publicly stated on July 22nd that Moonshot had built covert infrastructure to distill Anthropic’s Fable at scale, cycling through access methods to avoid detection, while developing K3. It’s worth noting that dubious methods were used in the past by the US to gain access to British industrial secrets. That history complicates the moral clarity of their position. The reality is that K3’s performance reflects a mix of distillation, engineering improvement, and scaling. Responses from Chinese stakeholders strongly emphasize independent development.

The Cost vs Efficiency Trade-Off

K3 brings cost dynamics into sharper focus. According to Benedict Evans, writing just before K3’s release on the broader question of token pricing:

“there are only 2 things you can say with certainty about token prices: we’re in a supply crunch, and this is unstable... the question at the end of that remains whether the foundation models have sustainable pricing power, strategic leverage and value capture, or whether they become low-margin commodity infrastructure providers. At the moment, I think every dynamic we can see points to the latter.”

A multi-polar model landscape is precisely the kind of development that strengthens Evans’s second scenario. And indeed, a recent Goldman Sachs research paper entitled “Navigating China AI models” anchored China’s top-performing coding models at roughly $1–2 per million tokens against $4–8 for their leading US rivals which is a 2-8x sticker-price gap. Very significant particularly for businesses with heavy agentic workloads.

Chinese open-weight models are certainly marketed as cost-efficient alternatives but some early analysis suggests a more complicated situation in practice. Independent testing points to relatively high token usage and slower responses on some tasks, while technical assessments indicate that self-hosting remains a production-scale undertaking with substantial memory and accelerator demands. Ben Thompson makes the version of this point asserting that tokens aren’t a commodity, because different models need different numbers of them to reach the same answer. Kimi reportedly needs nearly twice as many tokens as OpenAI's GPT-5.6 Sol to reach an equivalent answer according to Artificial Analysis-based benchmarking. If true, that narrows K3's roughly 2x per-token list-price advantage down to something closer to a sub-10% edge once real per-task cost is measured.

Headline pricing is therefore less useful as a comparator than effective cost, namely the resources required to complete real work at acceptable speed and quality. For enterprise buyers, the practical question is whether these Chinese open-weight models deliver enough capability, reliably enough, to justify adoption across common workloads to meaningfully reduce costs. Increasingly, the answer appears to be yes. They are good enough across a widening range of coding and knowledge-work use cases, which is exactly why adoption and evaluation are spreading even as questions on cost and efficiency remain open. OpenRouter data on comparing model tokens through H1 2026 shows a clear trend through the year:

One estimate goes further putting the share of startups already using Chinese models as high as 80%, a figure striking enough on its own to explain why frontier labs are currently lobbying so hard for protection. The Goldman Sachs research referenced earlier suggests that across general, agent, code, and data tasks on OpenRouter, Chinese models captured only 5–16% of dollar spend but 56–89% of token volume in the week of July 1 in the US, evidence that aggressive pricing, not raw capability alone, is what's shifting volume toward these models.

Routing as a COGS reduction initiative

Another trend is that as agentic workloads scale, token spend is starting to look like a cost line that businesses need to actively manage, the way they have for years with customer acquisition cost and cloud spend. Businesses are converging on routing to the cheapest model that clears the quality bar, and defaulting to open-weight where it’s good enough. The numbers already show it taking shape. Per-token inference pricing for flagship LLMs has fallen by roughly an order of magnitude since 2023 while request sizes have increased driven by reasoning tasks that increasingly dominate the total:

This is a pattern Kearney has begun formalising into a five-lever "AI FinOps" cost-management framework. DoorDash is a concrete case in point. It built a router that sends its hardest tasks to frontier models like Fable 5 and diverts everything else to cheaper open-weight Chinese models, according to Deutsche Bank research.

Converging towards commodity status or not?

The trajectory points to models becoming a commodity. Capability gaps narrow with each release cycle, and performance leadership now varies by task rather than sitting with one provider. The result is a multi-model ecosystem with parallel centres of development in the US, China and the EU, each with different strengths and optimisation strategies.

As capability converges, the basis of competition shifts. Evals still matter, but data, distribution, integration and trust become the real differentiators.

Security is where this shows up most sharply, and neither model type is simply safer. Open weights let anyone audit a model, and equally let anyone strip its safety training, including for offensive cyber use. Closed models have the opposite problem, with guardrails that can block legitimate defenders. In July, Hugging Face’s security team, blocked by US model guardrails mid breach investigation, turned to a Chinese open-weight model to analyse the attack instead.

Surveillance is a related worry, and how you view it depends on how you run the model. Calling K3 through Moonshot’s API means sending your data to a Chinese company, with all the jurisdictional questions that raises. Downloading the open weights and running them on your own hardware is a different proposition. You own the deployment, you control the network, and you can watch for anything trying to phone home. That doesn’t settle everything, since a model’s behaviour can still be biased in ways that are harder to detect, but it does separate the data risk from the model risk. Open weights are the version of this you can actually inspect.

So if capability is heading towards commodity, security and freedom from government interference are what buyers will still pay a premium for. That ambiguity is also why Washington is reportedly leaning towards highlighting security gaps in Chinese models rather than banning them outright. For a large corporation, that premium is easy to justify.

What This Means for US AI Companies

AI & Crypto czar David Sacks called K3’s benchmark result “concerning” on July 17, then argued for faster US innovation rather than restriction, while American digital news website Axios reported the same week that officials were weighing procurement rules and pressure campaigns to discourage Chinese-model adoption, short of an outright ban.

We see 3 possible options or reactions from the US government:

1. Restrict Chinese models. This would mirror the way China restricts Google and Meta today. However, Chinese models are free, open-weight, and already downloadable, so restriction mostly just raises costs for the US businesses that lose access. Chamath Palihapitiya suggests a ban would be a self-imposed price control rather than a security win.

2. Do nothing. Letting the market decide is risky for OpenAI’s and Anthropic’s IPO valuations if free/cheaper, capable alternatives are available for a growing share of work, which is exactly why both companies are now lobbying Washington for protection. The Wall Street Journal reported on July 21 that they’re warning of a “dystopian” AI future without regulation, with OpenAI’s Dean Ball calling an open-weight-dominated market a drift toward “AI communism.” Similar concerns were expressed when DeepSeek v1 launched. In that case, however, the initial concern was followed by normalisation and ongoing growth for closed models. We believe the market will bifurcate in a similar way to what happened in mobile with iPhone vs Android with closed models capturing most of the revenue and open weight models most of the tokens.

3. Subsidise. Matching Chinese state support with a US industrial policy of its own using cheap financing, tax breaks, or a direct backstop for the compute buildout behind OpenAI and Anthropic. The problem here is that federal finances are already stretched, and a subsidy large enough to offset the token cost gap is an open-ended commitment. It is, nevertheless, the approach Wall Street appears to believe the government will adopt. Jim Cramer, who in 2016 dismissed corporate fears of protectionism (which we have seen with Trump Tariff approach), is now openly asking Washington to protect OpenAI and Anthropic from Chinese competition. But is the US AI sector critical enough to eventually get bailed out the way banks were in 2008?

Underneath all 3 options sits a question about how much of the US AI buildout rests on debt. Credit default swaps on Oracle hit an all time high in July 2026, with Microsoft's not far behind, as bond market anxiety over AI capital spending spreads across the hyperscalers. Whether cheap, capable open weight AI poses a risk to AI related debt is contested, but several analysts argue markets haven't priced it in yet.

Analyst Luke Gromen of FFTT suggests the current situation represents a form of zugzwang (the chess position where you are obliged to move and every available move worsens your position). Restricting cuts smaller US firms off from cheap models. Doing nothing risks a disorderly unwind of debt-financed AI valuations. Subsidising adds to inflation that’s already elevated. It explains why “wait and see, but test the models seriously” is emerging as the default posture, rather than an actual strategy.

Implications For Leaders

The implications will vary depending on what constituency they represent:

• Technology leaders: Adopt a multi-model sourcing strategy over the next 12–24 months. Evaluate models at the workload level and maintain flexibility to match providers to specific use cases using a model router architecture that allows you to easily switch between models and integrate new ones. Track both technical progress and regulatory conditions. Deployment feasibility will continue to vary by region and use case, shaping adoption pathways.

• Security Leaders: Treat open-weight, foreign-origin models as a new category of supply chain risk, not just a procurement choice. Evaluate their provenance with at least the rigour applied to closed-API vendor reviews, arguably more, since there’s no vendor left to hold accountable.

• Financial Leaders: Base decisions on observed performance in realistic conditions. Include token efficiency, latency, and infrastructure requirements in total cost assessments. Closely track and monitor actual token usage.

• Investors: Assess where durable value accumulates as capability converges. Platforms, applications, and integrated solutions may capture a growing share of value relative to standalone models which are starting to resemble airlines as capital intensive, thin margin businesses. Would you invest in OpenAI IPO given that dynamic?

Moonshot released Kimi K3 open weights on 27th July. The model has already been integrated and made available by DigitalOcean (a developer-focused cloud platform). Their token charge for serverless inference for Kim K3 and its 2.6 predecessor are as follows:

Ollama are also offering Kimi K3 for the same price. One which is roughly 30% the list price of using Fable today:

It will be fascinating to see whether it holds up under independent testing, which of restrict, subsidize, or do-nothing the administration actually commits to, and whether OpenAI’s and Anthropic’s planned IPOs stay on track.

Interesting times!

Epilogue

We asked Kimi K3 to review this article. Self-assessment has obvious limits, but here is what it said:

The article is largely accurate on its core factual claims, which check out against multiple independent sources: Kimi K3 did launch in mid-July 2026 as a 2.8 trillion-parameter MoE model, topped Arena‘s Frontend Code Arena ahead of Anthropic’s Fable 5, and released its full open weights on July 27. The reported reactions from David Sacks, Michael Kratsios, and Dean Ball are all verified, as are Microsoft’s evaluation of K3 for Azure/Copilot and Benedict Evans’s token-pricing analysis. Where the piece becomes less reliable is in blending these verified facts with speculative analysis. The distillation allegation is handled with appropriate caveats, though the article’s strategic analysis leans toward treating it as substantially true despite noting it remains unproven. Overall, it’s a well-sourced snapshot of real events, but readers should treat its forward-looking policy and market conclusions as informed opinion rather than established fact.

Apple just sued its former Siri integration partner OpenAI over alleged trade secrets, while Beijing blocked Meta from buying the hottest agent startup and handed it to Tencent instead. Plus some charts that will make you go WOW. Probably nothing you don’t already know, but the sheer size of some of this AI stuff is still incredible.

CxAI team

1. SK Hynix raises $26.5Bin the largest foreign IPO in US history

The South Korean memory chipmaker’s New York listing reflects desperate Western demand for High-Bandwidth Memory (HBM), the critical component Nvidia needs for its AI accelerators. The capacity to train and run frontier models now depends on a handful of Asian semiconductor firms, see below our Chip recipe chart. The irony is that American AI dominance now hinges on persuading a Korean company to do what Intel could not (ie to become the global chipmaker). TechCrunch

2. Apple sues OpenAI over alleged trade secret theft

Apple filed suit against OpenAI and its Chief Hardware Officer Tang Tan of misappropriating IP related to smartphone and wearable design. Tan spent 24 years at Apple, most recently as VP of product design for the iPhone and Apple Watch, before joining OpenAI. The lawsuit suggests OpenAI’s first hardware device could be a smartphone, and 1 year after OpenAI acquired Jony Ive’s device startup io for $6.5B. This signals the stake in the arrival of a new AI native device that could change the entire Apps ecosystem. TechCrunch Axios. AP News. dw.com

3. Beijing blocks Meta’s $2B acquisition of Manus

Chinese regulators intervened on national security grounds, forcing Meta to unwind its deal for the AI agent company and leaving Tencent to negotiate a majority stake instead. The surprise is that Meta got as far as signing a deal before Beijing acted, suggesting regulatory lines remain blurry until they are enforced. The Decoder. CNN. CNBC. npr.org

4. OpenAI launches GPT-5.6 family with 3-tier pricing

Does this suggest a price war Open AI <> Anthropic coming soon? The Decoder TechCrunch. en.wikipedia.org. community.openai.com

5. China enforces the world’s first dedicated rules banning emotional manipulation by AI companions.

The new framework targets romantic and parasocial chatbots, prohibiting features designed to foster user dependency or simulate intimate relationships due to mental health risks and the potential for foreign exploitation of personal data. For companies operating in this space this is a massive watch out. AI News CNBC. aisafetychina.substack.com. techpolicy.press

Chart(s) of the Week

This week we have plenty. Worth mapping who does what in the chip world, so we don’t lose sight of it plus some others interesting charts for this months Deutsche Bank July deck that will make you WOW:

How to “cook” a chip, not that straight forward with 7 countries involved

On Frontier models market share: Claude gaining share fast, now close to 30%

Private AI Investment in the US is HUGE compared to the rest of the world. However it would be very interesting to know how large is Chinese government investment in the space

But are they spending a bit too much, see below the largest 5 US Hyperscalers spending vs. their cash-flow

Whilst, valuations in this space are getting bananas, not just for typical guys (OpenAI, Anthropic) Micron & SK Hynux are both memory chips manufactures

CxAI Weekly Pulse · 2026-07-13 · AI news for senior leaders

AI’s speed cut both ways this past week. Ford rehired around 350 veteran engineers after its AI quality systems missed defects instead of catching them, and topped JD Power’s mainstream-brand quality rankings for the first time in 16 years. Security forecasters now expect around 66k software vulnerabilities this year, mostly because AI agents are finding flaws faster than humans ever could. And the community talked about two things. First, OpenAI reportedly offering the US government a 5% stake worth around $42.6B, and then Jodie Foster saying Brad Pitt’s F1 felt “made by AI”. Machines are getting faster at everything. The question is whether we are retaining the judgement to steer them.

CxAI team

1. Microsoft commits $2.5B to embed 6K AI engineers

The company launched “Frontier Company,” a forward-deployed engineering unit designed to integrate AI directly into customer operations rather than selling standalone software. This reflects a broader industry shift from product sales to outcome-based partnerships, where success is measured by workflow transformation, not licence revenue. This signals that AI vendors now expect to sit inside your organisation, shaping processes from within. Amazon committed $1B to its own unit two days before Microsoft, and OpenAI’s $4B+ venture with TPG and Anthropic’s $1.5B venture with Goldman Sachs, Blackstone and Hellman & Friedman launched in May. THE DECODER . MICROSOFT

2. Security vulnerability reports surge as AI models hunt for bugs at scale

Forecasters now project 66k Common Vulnerabilities and Exposures disclosures this year, a historic high driven by autonomous agents systematically scanning codebases. The same capabilities accelerating discovery are also enabling attackers to generate malware that evades signature-based detection. Chief Information Security Officers face a paradox. Deploying AI to find flaws faster while simultaneously defending against AI-powered exploits. Investment in runtime monitoring and behavioural detection is now essential, not optional. The irony is that organisations delaying AI adoption may be more exposed, lacking the automated defences their competitors are building. THE DECODER

3. Ford rehires 350 veteran engineers after AI fails to catch design flaws.

The automaker reversed course when its automated quality systems amplified errors rather than detecting them, creating a feedback loop that veteran engineers had to untangle. This hybrid approach, pairing institutional expertise with automation, delivered hundreds of millions in cost savings once reinstated. For operations leaders, the lesson is stark: AI excels at pattern recognition but lacks the contextual judgment built over decades of hands-on experience. Expect similar reversals in other asset-intensive industries as companies discover the limits of full automation. The episode also suggests that the “AI replacing workers” narrative oversimplifies what is actually a rebalancing of roles. Further evidence of this from a survey of US companies suggesting that companies who invest in AI also add positions, but not immediately. TECHCRUNCH . TheRegister

4. SpaceX acquires Cursor for $60B, signalling consolidation in AI coding tools.

The deal brings one of the most popular AI-assisted development platforms under SpaceX’s umbrella, raising questions about whether Cursor will remain open to models from OpenAI and Anthropic. Enterprise customers relying on Cursor now face potential lock-in to a single ecosystem. The purchase also suggests SpaceX views software productivity as integral to its aerospace ambitions, not a side business. WIRED

5. Anthropic wins California government contract at half-price rates for Claude.

Governor Newsom’s administration secured discounted access to Claude for state agencies, establishing a precedent for public-sector AI procurement. The deal positions Anthropic as a favoured vendor for government use cases, potentially influencing federal procurement standards. Critics argue the discount could distort competition, favoring one provider over others. The state vs federal divergence is the more interesting angle. TECHCRUNCH

6. Coinbase cuts AI spending by nearly 50% through smarter routing

The crypto exchange migrated workloads to cheaper open-weight models from Chinese providers and implemented aggressive caching, proving that inference cost, not model quality, is the primary constraint. The shift validates a broader trend: enterprises can achieve comparable results at a fraction of the cost. Some projections suggest 80% percent of workloads will run on models 99% cheaper within 18 months. The move also raises questions about reliance on Chinese infrastructure amid ongoing export control debates. THE DECODER

Our take: The agents are fast, but they still need a boss

Our read: Ford’s AI amplified the very flaws it was supposed to catch.

This week crystallized a tension that will define enterprise AI for years: autonomous agents are delivering measurable productivity gains, yet the same speed that makes them valuable also makes them dangerous without human oversight.

The security data reinforces the point from a different angle. With 66K vulnerabilities projected this year, many discovered by autonomous agents, we are watching machines find problems faster than organizations can fix them. Attackers are using the same tools to generate malware that evades traditional defences. Runtime monitoring and behavioural detection are no longer optional investments. They are table stakes.

For senior leaders, the strategic question is no longer whether to deploy agents, but how to govern them. The market is rewarding speed, yet last week’s events show that speed without judgment creates new categories of risk. Expect the next quarter to bring sharper regulatory scrutiny and, inevitably, the first major incident where an unsupervised agent causes reputational or financial damage at scale.

Chart of the Week

With the AI ecosystem getting noisier by the week, here’s our simple map of who does what as a good reminder. Solid colors here mark the category leaders.

Source: Bessemer, Intel Capital, Stackone

CxAI Weekly Pulse · 2026-07-04 · AI news for senior leaders

A contradiction about the AI industry is that the companies warning us loudest about its dangers are the same ones racing to build the most powerful systems. Anthropic now sits at nearly a $1T valuation while arguing only its dominance can keep us safe. Meanwhile, the White House is asking OpenAI to pump the brakes on GPT 5.6, just weeks after a separate export control order forced Anthropic to pull its own most advanced models, Mythos 5 and Fable 5, from the market entirely.

CxAI team

1. White House asks OpenAI to limit GPT 5.6 release to select partners

The Trump administration wants OpenAI’s newest model (GPT 5.6) to be shared only with a select group of partners. This is driven by safety concerns that echo growing anxiety about autonomous capabilities outpacing oversight. The administration see GPT‑5.6 as a potential frontier cyber weapon and is worried about national‑security and cybersecurity misuse. The irony is sharp as the administration most hostile to regulation in sectors such as energy and infrastructure, data centre construction, environmental protection and nuclear power is now applying brakes to AI’s fastest-moving companies. TechCrunch Computing Guardian

2. Anthropic reaches nearly $1T valuation

Anthropic has been valued at almost $1T after 5 years of warning the world about the dangers of AI. Dario Amodei, its founder, left OpenAI in December 2020 to found an AI hyperscaler PBC (Public Benefit Company) focused on proper safety and privacy guardrails. Anthropic argues its own market dominance is the best path to safety, a claim that positions commercial success and existential caution as inseparable. The question now is whether concentrated power in a few safety-conscious firms actually reduces risk or merely concentrates it differently. The uncomfortable truth is that Anthropic’s business model requires it to build the very systems that it says could be catastrophic. Wired

3. Google rolls out default data harvesting for AI training, prompting opt-out concerns.

Google has enabled a new default privacy setting, “Search Services History,” which automatically collects user media including images, voice recordings, and file uploads to train its AI models. The opt-out design places the burden on users to protect their data, raising questions about informed consent and data sovereignty. For organisations, this development underscores the need for board-level oversight of how employee and customer data may be used by third-party AI platforms. Regulators in Europe are likely to scrutinise this approach under existing frameworks. Google argues aggregated data improves products for everyone, but users have little visibility into what happens once their information enters the training pipeline. Wired

4. Anthropic accuses Alibaba of pilfering Claude capabilities.

Anthropic has alleged that Alibaba copied Claude’s capabilities, escalating tensions between Western AI labs and Chinese technology giants. The accusation points to broader anxieties about intellectual property protection as model architectures and training techniques become the most valuable assets in the industry. The dispute may accelerate calls for international AI governance frameworks. Yet the deeper issue is that model capabilities are increasingly difficult to attribute or protect, making such allegations hard to prove and easy to weaponise. AI Business

If you are interested in what Anthropic claimed Alibaba did, see below:

• Anthropic says operators linked to Alibaba and its Qwen AI lab created ~25,000 fraudulent user accounts, which generated ~29 million interactions with Claude over a short period

• Those interactions weren’t generic chatbot use; they allegedly focused on Claude’s frontier capabilities: advanced software engineering, agentic reasoning, long‑horizon task handling, and strategies for complex decision‑making.

• Anthropic believes this is a “distillation attack”: using the outputs of a powerful, proprietary model as training data to “distill” its behaviours into a cheaper or smaller rival model.

Worth mentioning that Alibaba denies any wrongdoing.

Our take: The safety paradox is now the industry’s defining tension.

“Anthropic’s business model requires building the very systems it says could be catastrophic.”

The most striking pattern this week is not any single deal or product launch. It is the emergence of a fundamental contradiction at the heart of the AI industry. The companies most vocal about existential risk are also the ones racing hardest to accumulate power, capital, and capability. Anthropic now argues that its own market dominance is a prerequisite for safety, while simultaneously warning that advanced AI could cause mass destruction. OpenAI is building custom chips to scale faster, even as the White House asks it to slow down.

This is inevitable in a winner-take-most market where ceding ground means ceding influence over how the technology develops. Yet the logic creates a troubling dynamic. Safety rhetoric becomes a competitive moat, not a shared commitment. The EU AI Act was designed for assistive tools, not for autonomous agents that execute decisions in the real world. Governance frameworks are already outdated.

For leaders, the practical implication is clear. You cannot outsource AI risk management to the labs that build the models or the regulators who lag behind them. Watch how the GPT 5.6 restriction plays out. If governments begin gatekeeping frontier releases, the rules of engagement for enterprise AI adoption will change faster than most organisations are prepared for.

Chart of the Week

Sources: Reuters (Series D in 2023, 2024 Amazon investment, 2025 Series F, 2026 Series F)

The week’s sharpest contradiction! Governments are simultaneously demanding that AI companies prove their models cannot be hacked, while security researchers say that goal may be technically impossible. Meanwhile, KPMG fabricated case studies to sell AI adoption projects.

1. White House demands Anthropic block all jailbreaks, but experts say that may not be possible

The Trump administration just blocked foreign access to Anthropic’s two most powerful models (Fable 5 and Mythos 5) after researchers discovered a way to hack them to enable cyberattacks. Amazon researchers found it, Andy Jassy called the White House, and the rest is history.
Worth noting: these models are only accessible to around 200 companies (a significant number of them foreign) so the immediate impact is limited. But the signal to the industry is loud - preventing this kind of hack is now a national security issue. If widely applied, compliance costs will rise for something that is, frankly, close to impossible to fully guarantee.
In the meantime, we wonder, why does Anthropic keep ending up on the Trump administration’s naughty step? Could the reason be that Anthropic refused to let the military use its models for autonomous weapons? It would have been hard to believe a few years ago, like many things today. WIRED

2. KPMG fabricated AI case studies to sell AI adoption

The consulting firm reportedly invented examples of successful AI implementations in a client-facing document intended to encourage enterprise adoption, raising serious questions about trust and verification in professional services. The revelation comes at a moment when enterprises are under pressure to prove measurable ROI before scaling AI investments. This is a stark reminder that 3rd-party claims about AI value require validation and scrutiny. The uncomfortable truth, however, is that fabricated case studies are easier to catch than subtler forms of AI-driven misinformation embedded in everyday business tools. The Decoder

3. SpaceX valuation reaches $2.6 trillion, briefly surpassing Amazon.

SpaceX market capitalization has ballooned to historic levels, driven by investor enthusiasm for its launch business, Starlink connectivity, and positioning as critical infrastructure for AI and defense applications. The valuation surge reflects broader appetite for companies perceived as essential to the next phase of AI buildout, particularly those with differentiated access to compute, connectivity, and physical infrastructure. Watch for spillover effects on valuations for data center operators, power suppliers, and satellite communications. TechCrunch

4. EU publishes AI content labelling playbook ahead of the AI Act’s August deadline.

The European Union has released its official code of practice for AI-generated content, outlining transparency requirements that will apply to platforms and providers operating in the bloc. The playbook specifies disclosure obligations for synthetic media, watermarking expectations, and record-keeping standards intended to make AI outputs traceable. For businesses deploying generative AI in marketing, media, or customer-facing applications, compliance planning must now move from policy review to operational implementation. The practical challenge is that enforcement mechanisms remain unclear, and cross-border consistency is not guaranteed. The question to watch: whether this framework becomes a global template or remains a European outlier. AI News

Our take: Governance Is the New Growth Constraint

“The White House is demanding outcomes the technology cannot deliver.”

This week made one thing clear: governance is no longer a compliance checkbox. It is the binding constraint on how fast and how far AI can scale. The U.S. government is treating frontier models as strategic national infrastructure, demanding that Anthropic prove its systems cannot be exploited, even as technical experts say that standard is unachievable. In Europe, the AI Act’s content labelling playbook is moving from policy to operational requirement, forcing every company deploying generative AI to rethink disclosure, watermarking, and record-keeping before August.

For business leaders, the implication is practical: your AI roadmap now depends on your governance posture. If you cannot demonstrate transparency, provenance, and human oversight, you will not get permission to deploy in regulated markets or high-trust workflows. The KPMG fabrication story is a warning that even advisory credibility is under scrutiny. Meanwhile, the capital markets are rewarding infrastructure and inference at historic valuations, betting that AI will embed everywhere, but that bet only pays off if governance frameworks catch up.

Watch for the gap between what regulators demand and what vendors can deliver. That gap will define which deployments scale and which stall.

Chart of the Week

The US government ordered Anthropic to take its most advanced models offline. Meanwhile, Jeff Bezos’s new venture raised $12B at a $41B valuation to build robots for the physical world. The lines between AI policy, capital deployment, and competitive positioning are blurring faster than most organisations can adapt.

1. US gov. orders Anthropic to disable most powerful models

Anthropic has taken Claude Fable 5 and Claude Mythos 5 offline exactly 72 hours after its launch due to national security concerns from the US Government claiming someone found a way to bypass its safety guardrails. The order targeted foreign nationals (easier to implement than for American citizens) however Anthropic chose to shut down access for all customers rather than implement selective restrictions. TechCrunch

2. OpenAI files for IPO and plans to rebuild ChatGPT

OpenAI has filed for a public offering and signalled that it expects to go public within the next year. At the same time, leadership has described a strategic shift away from chat toward an agent-based architecture that integrates coding tools, image generation, and external partner services. Watch for pricing changes and new enterprise tiers as OpenAI seeks to demonstrate durable monetisation. The Decoder

3. UK bets on AI supercomputer to reduce US dependence

The UK government is investing in national AI supercomputing, advanced chip procurement, and ecosystem support as part of a sovereignty-driven strategy to build domestic AI capacity. The £1.1 billion AI Hardware Plan reflects growing political concern about reliance on foreign infrastructure and the desire to capture more of AI’s economic value locally. National AI projects have historically struggled to match the pace and scale of private hyperscalers, and the UK may find itself building infrastructure that is outdated by the time it is operational. Wired

4. Jeff Bezos’s Prometheus raises $12B

Prometheus has closed a $12B round at a $41B valuation, positioning itself as the leading investor-backed effort to bring AI into engineering, manufacturing, and drug design. The company is pursuing “physical AI”: systems that can reason about and manipulate the real world rather than just generate text or images. Hardware-AI integration remains far harder to scale than software, and Prometheus will face supply chain, regulatory, and workforce challenges that pure-model companies can sidestep. TechCrunch, ArsTechnica

Our take: Government is now a Variable in AI Strategy

“Anthropic’s own safety warnings about its models’ capabilities may have accelerated the very regulatory action that now constrains its commercial reach.”

This week made one thing clear: AI is no longer just a technology decision or a procurement choice. It is a strategic asset, and governments are treating it accordingly. The US order to disable Anthropic’s most advanced models was not a distant regulatory development. It was a real-time disruption to enterprises relying on frontier capabilities.

For senior leaders, the implication is operational, not theoretical. Vendor concentration risk is no longer about service-level agreements and uptime. It is about whether your model provider can be switched off by executive order. Multi-model architectures, fallback routing, and geographic diversification are moving from nice-to-have to essential. At the same time, the capital intensity of AI infrastructure, visible in Amazon’s $17.5 billion borrowing and Prometheus’s $12 billion raise, means the winners will be those with both financial scale and regulatory resilience.

Chart of the Week

Just to be clear, the daily reports are entirely generated by AI. Even so, we have been doing this a while and have numerous quality gates and checks on content. This weekly summary is also AI-generated, but curated and edited by us to surface the stories and themes that matter most to senior leaders.

What We Learned This Week

The Trump administration is exploring taking an equity stake in OpenAI, which would mark the 1st direct US government ownership position in a frontier AI company. Meanwhile, Uber burned through its entire AI budget in 4 months and had to cap employee spending, a reminder that AI costs scale faster than most finance teams expect.

1. Trump administration considers taking equity stake in OpenAI

The White House is exploring whether the federal government should acquire a direct ownership position in OpenAI, potentially through a new Public Wealth Fund that would let the American public share in AI’s financial upside. Federal govt taking an ownership stake in a private company is essentially a form of partial nationalisation which also happened with a 10% stake in Intel last year. Watch for whether this approach spreads to other AI companies though as the FT put it: “There is no precedent for the government taking a stake in loss making AI labs collectively worth trillions of dollars”. This happens at the same time as Bernie Sanders proposes a one-time tax on AI companies in the form of stock, just 50% of it all. Techcrunch.com FT

2. Alphabet plans record $85B raise to fund AI infrastructure expansion

Google’s parent company is preparing the largest capital raise in tech history to finance data centres, chips, and cloud capacity for its AI business. This move underscores that in AI competition balance sheet strength may matter more than model novelty. It also reveals how dependent even the richest tech firms have become on external financing to sustain AI growth. As a reminder, total Tech industry AI Capex spending amounts to $700B, not a small feast. *Techcrunch.com Yahoo Bloomberg*

3. Uber caps employee AI spending after blowing through budget

The ride-hailing company had to impose spending limits ($1500/month) after employees exhausted the full-year AI budget by April. This is a warning that AI adoption curves can become really expensive, especially when it is really hard to assess its impact. In Uber’s case the overspend appears to have been driven primarily by agentic AI coding tools and not customer support systems. This problem will likely get worse as hyperscalers which are operating at cost increase their prices. *TechCrunch Bloomberg*

4. Google DeepMind’s Gemma 4 12B brings multimodal AI to laptops with 16GB of RAM.

Open-weight models, where the weights are available but not the actual source code, can handle multimodal input in an architecture that is able to run on consumer hardware (such as a MacBook Pro). This lowers the barrier for developers who want to experiment without paying enterprises, it suggests a future where edge (on device) deployment is viable for internal tools and privacy-sensitive applications. The downside of Gemma 4, is that it may not be as potent as cloud based models. The Decoder

Our Take: Watch out for AI pricing evolution

This week made one thing clear: the AI industry is done pretending that growth alone will pay the bills. Meta is testing a $200-per-month consumer agentic AI product a sharp departure from its advertising-funded business model. Anthropic is preparing for public-market scrutiny through an IPO, and Alphabet is raising $85 billion because even its cash pile cannot keep pace with infrastructure demand. It is a structural response to the reality that AI at scale is extraordinarily expensive to run.

The implications are equally clear. Vendors will increasingly price AI by outcome, workflow, or seat rather than by token, and the economics of adoption will tighten. Uber’s cautionary tale is a reminder that finance teams should expect AI spending to behave more like cloud computing than like software licensing, with similar needs for usage governance and capacity planning.

The deeper question is whether subscription models can sustain the capital intensity of frontier AI. Anthropic and OpenAI are betting that enterprise customers will pay premium prices for trust, safety, and integration. The next few quarters will test whether that bet holds, especially as open-weight models like Gemma 4 lower the cost of experimentation. Watch for how pricing evolves once the IPO window opens.

Chart of the Week

CxAI Weekly Pulse · 2026-06-07 · AI news for senior leaders

This week marks one year since our first post, so we wanted to cover a topic we have been sitting on for a long time, personalisation. It is a challenge we ran into repeatedly at Depop, Gumtree and Monzo. We will explore why retailers have struggled to deliver truly personalised commerce, why recommendation engines are not really personalisation, and why LLMs and agentic interfaces may finally change the equation.

We wrote this with John Cragg, former Head of Data at Depop, Reveri’s CTO and currently Director of Data & AI at Swap Commerce.

Recommendation is not personalisation

These two terms get used interchangeably and mean different things as illustrated below:

Recommendation is a matching function which is tied to the product. Given a product, rank others most likely to be relevant.

Personalisation tailors an experience to a specific person across every touchpoint, including search. Spotify’s “Good morning, Maria” or StitchFix’s “Style Shuffle” are the right shape by being person-centric.

A retailer can have a sophisticated recommendation engine and still not be doing personalisation at all. They rank products well, but they do not rank them for you. This post is about that second, harder problem.

The Amazon myth

Amazon is treated as the gold standard, but what it actually has is a world-class recommendation engine. It is excellent at predicting replenishment and surfacing accessories for items you just bought. It is weaker at understanding who you are as a person, what you aspire to wear, or what trend is forming in your social orbit. For commodities, that doesn’t matter so much. For fashion, home and other considered purchases, those limitations become far more obvious.

The result is a retail industry that has largely confused segmentation with personalisation.

Researching this post, we discovered that Amazon does have a personalisation settings panel tucked inside your account (image 1). Once you find it, the Top picks it serves up tell their own story: in our case, vodka, Pepsi, NFL and NBA merchandise (Image 3). We live in London, have never bought any of it, and certainly not the American version.

Amazon is not blind to the gap. Its COSMO system uses LLMs to read intent beyond simple keywords, a real step forward, but it is still anchored to purchase patterns rather than who you are.

Who is actually doing this well?

Despite being largely unknown in the West, Chinese platforms are clearly leading, and the reason may not be what you expect.

Taobao, Alibaba’s consumer marketplace, operates one of the most complete real-time personalisation loops in global retail.

The platform continuously reshapes each user’s storefront using signals like scroll velocity, dwell time, and interaction patterns, while its merchant ecosystem responds almost instantly with changes to pricing, inventory, and creative, compressing the lag between “interest detected” and “product available” to near zero.

What makes the engine truly unique, though, is that Alibaba is fully integrated across the rest of the customer’s digital life. It owns the payments (Alipay, akin to PayPal or Apple Pay), the location data (AutoNavi, China’s Google Maps), the mobile browser (UCWeb, like Chrome), the video platform (Youku, China’s YouTube), and the logistics network (Cainiao, comparable to FedEx) all resolved to one customer profile under a single consent. In the West, those 5 categories sit inside five different companies.

TikTok Shop is another strong example. Global GMV reached ~$33bn in 2024 and is on track to roughly double in 2025 (Statista). Every like, comment and second of watch time is a data point, generated millions of times a day. Four things made it work where Instagram’s commerce push did not (Instagram pulled its Shop tab in early 2023 and shut down live shopping that March):

1. TikTok Shop has a dedicated tab, but the feed’s behavioural data still powers the shopping experience.

2. The recommendation engine updates every minute.

3. A single model predicts likes, watch time, comments, shares, add-to-cart and purchase.

4. The entire checkout experience happens in TikTok.

Zalando is the most serious European attempt. Its AI-powered discovery feed is a real shift from search-first platform to fashion destination, with a 13% lift in items added to bag and an 8% drop in size-related returns reported in its 2025 full-year results.

That said, moving users from social platforms to a retailer’s own app for discovery is uphill. As many consumer CEOs know, you need to meet customers where they are, rather than pulling them to where you are. From our testing, the onboarding flow still feels heavy, and the AI assistant loses coherence after a few interactions. It is still in beta, so there is room to grow.

Three main ingredients

Truly personalised commerce depends on three inputs rarely available inside the same retailer stack at the same time: user data, product catalogue and live social trends:

1. User data

Not purchase history in isolation, but demographics, life stage, offline behaviour, loyalty tier, social signals, basket abandonment, dwell time and search intent, all joined to a single customer identity across every channel. Without this, the engine produces absurdities such as a customer who just bought a coat being emailed the same coat the next day.

The tools to support user data acquisition have existed for years: Customer Data Platforms (CDPs) and identity resolution engines like Segment, and mParticle. The missing piece is the data foundation.

Where agentic changes the picture. Most online retail traffic is guest, unauthenticated, or first-visit. To personalise for that user, the retailer has to bootstrap a profile from scratch, usually via a long onboarding questionnaire. Each extra question adds friction; each skipped question weakens the signal. Zalando’s brand-preference flow, which we tested ourselves, is the canonical example. By the time you finish picking your brands, you have already lost patience.

Agentic shopfronts change the unit economics of that exchange. Replace a 50 question flow with a chat box: “Tell me what you’re looking for and why.” A user replies as they would to a personal stylist: “A dress for a festival in late September. I’m a size 12. Here are three photos of outfits I like.” In thirty seconds the retailer captures richer preference data than most onboarding flows collect in minutes.

Another foundational layer is zero party data: information customers intentionally choose to share with retailers. The challenge is incentive alignment. Users only share meaningful data when they receive meaningful utility in return: better recommendations, more accurate sizing, earlier access or reduced search friction.

In theory, retailers could also import parts of a user’s social graph and preference data directly from external platforms, subject to consent and portability regulation. But turning fragmented social data into something operationally useful remains extremely difficult. Tools like Onfabric are beginning to explore this layer, though adoption remains limited.

2. Catalogue depth and AI-enriched attributes

Personalisation is a long-tail problem. Below roughly 10,000 SKUs, the maths does not permit meaningful differentiation: everyone sees the same products. But depth alone is not enough. The catalogue also needs rich semantic attributes (style, occasion, trend, fit, fabric) so that user preferences can be matched to products. AI-driven enrichment tools generate these from images and descriptions, turning a flat inventory into a navigable space. Without enrichment, even deep catalogues default to surfacing the same bestsellers for everyone.

3. Live trend signal

Fashion is emotional. Nobody buys a blazer because a regression model said so. They buy it because they saw it on someone they admire. Until TikTok and Instagram became dense enough to mine at scale, that signal was invisible to retailers, who were attempting emotional commerce with purely transactional data and wondering why it felt cold. Social listening platforms (Heuritech, Pulsar, Trendalytics) now extract actionable trend signals in near real time. Wiring them into the recommendation layer is engineering work.

A fourth enabler deserves a mention: the modelling infrastructure. Classic collaborative filtering could not handle the combinatorial explosion of “this user × this catalogue × this trend × this moment.” Embeddings and LLM-enhanced retrieval can.

Until recently, the computational architecture required to combine user data, product catalogues and live trend signals simply did not exist at production scale.

In Summary

Truly personalised commerce remains an incredibly difficult problem to solve and we are still years away from most retailers delivering experiences that genuinely feel personal.

One path is that retailers finally get their data, catalogues and architecture in order, combining the three key ingredients outlined above to build highly adaptive recommendation and discovery systems, closer to what we already see in parts of China.

The other possibility is that agentic commerce changes the equation entirely. Instead of retailers trying to infer intent from fragmented signals, users may increasingly interact directly with AI agents, expressing what they want conversationally. Those interactions could generate far richer preference data while also bypassing many of the cold-start and identity challenges retailers struggle with today.Most likely, the future will involve some combination of both..

Either way, the underlying problem remains the same: retailers have historically lacked a deep understanding of the customer in front of them. For the first time, the technology foundations required to solve that problem may finally be emerging.

Time will tell. In the meantime, we would love to hear your thoughts, especially if you know of businesses pushing the frontier forward in this space.

We hope this was useful.

Very best,

The CxAI team

We’re both jealous and curious. How has the shift really felt? What’s actually changed day-to-day? And through our Consumer AI lens, we want to explore what this means for other executives considering a similar path.

With AI, starting something ambitious seems easier or at least cheaper than ever before. But is it as rosy as it sounds? What are the real challenges beyond the obvious ones? What should you actually know before making the leap?

Over to you, Kim!

Kim’s Journey

Last year I was MD at Gumtree, leading business intelligence, marketing, product strategy, and customer experience. Before that, I held C-level and board roles at Checkatrade and giffgaff. 20 years building and leading product, marketing, commercial, and data teams at scale.

My calendar was 80% meetings. Product strategy reviews. Roadmap prioritisation. Quarterly business reviews. One-to-ones. The value I added was main judgement and decision-making. My job was to make sure the people who did had everything they needed

Then, suddenly: silence.

No Slack notifications. No direct reports. No stakeholder alignment meetings. No board packs. Just time, and nobody to delegate to.

Note: AI-generated illustration (Flux LoRA, trained on author’s likeness)

My first instinct was to hire and build a small team (a data engineer, a data analyst). Every senior role I’d held had taught me the same: if you want to build something, hire the people who can build it, and spend your time making sure they have what they need.

But facing another round of hiring meant rebuilding the delegation layer I had just left. I had to ask whether the answer I had learned over twenty years was still the right one for the tools that existed now.

Today I run Early Phoenix, an AI-powered business intelligence platform that tracks pricing, promotions, and competitive moves across European supermarkets for FMCG brands. Every day, our clients wake up to an alert telling them exactly what changed overnight. I built it entirely through AI. No team. No engineers. No analysts. Just me, a microphone, and Claude Code.

The question I get asked a lot is: “what has actually changed in my day-to-day?” The short answer is that the tools changed completely and the judgement did not.

What 20 years actually gave me

I expected the transition to feel like taking old skills and applying them to a new kind of team. Management experience in, agent productivity out. That is the clean version of the story. It is not quite the truth.

Some skills do carry over:

• Writing a clear brief. An agent with a vague brief produces vague work, exactly like a person would.

• Setting context without overloading. Too much context and the agent loses focus, too little and it guesses.

• Defining what good looks like. If you cannot describe the output you want, you will not recognise it when you get it.

• Trusting but verifying. The default posture with any new hire, and the permanent posture with any agent.

If you have spent years managing people, these muscles are already built.

But the part that transfers most is harder to explain.

Last week, an agent delivered a pricing dashboard that looked clean at first glance. Well formatted, numbers in the right ranges, comparisons laid out logically. But I had seen enough dashboards over twenty years to know the comparison methodology was wrong. That instinct, the feeling that something is off before you can articulate exactly why, comes from having seen what good looks like hundreds of times and having made a lot of mistakes along the way.

That pattern recognition is not something you can prompt your way to. It is the single most valuable thing I brought from my old career into this new one.

What Mollick and his co-authors call the Jagged Frontier (Co-Intelligence, 2024) is the honest version of what does not transfer. The way you work with AI is fundamentally different from managing people. You are always pushing to find the limits of what you can trust and what you cannot, and those limits move because the models keep getting more sophisticated, and the tools and infrastructure around them are shifting just as fast. You can switch between different models that give you confidence to build higher quality work over time. But understanding where human intervention is still required versus what can be handed off, that is a continuous exercise that never ends. There is no stable foundation to stand on. The foundation is moving.

How I Built 30 Agents

An agent, in my world, is not a piece of custom software. It is a Claude Code session running in its own terminal tab, pointed at a specific project directory, with a brief that tells it what to do. On disk, each agent has a state file that captures its current context, so it can be stopped and restarted without losing its thread. There is no custom code to write. Building an agent is closer to writing a job description than writing software. A simple one takes 30 minutes. A complex pipeline evolves over weeks, but the first useful version runs within a day.

Each one has its own identity, memory, and conversation history. Some build pricing dashboards for UK snacks or track competitive moves for soluble cocoa in Spain. One monitors the UK pet welfare market and has helped generate evidence that has reached Defra and welfare organisations. Others handle my email, prepare me for meetings, and draft my communications.

Source: AI-generated illustration (Flux LoRA, trained on author’s likeness)

To make that concrete: the pet welfare pipeline is the clearest example. A set of agents collect live pet sale listings from 8 platforms across the UK. Prices, seller details, health documentation, breeding information. A validation agent checks the data against live counts and field coverage thresholds. If it passes, a builder merges everything into a standardised dataset. An analytics agent generates month-on-month comparisons: which platforms are improving welfare disclosure, which are getting worse, where unlicensed breeding patterns are emerging. That output feeds directly into evidence packs for Defra and welfare organisations. No human touches the data between collection and insight. I review the final output and decide what to share.

The stack that runs all of this is deliberately simple.

The whole operation costs a fraction of one junior hire’s daily rate.

What My Day Actually Looks Like

Here is the day-to-day. The principles have been written about plenty; the specifics of how one person actually runs 30 agents have not.

The structure of the day flipped. Before, I would try to squeeze focus work in before 9am, then meetings ran until 5, then evenings were spent catching up on email and Slack. Now, mornings are five hours of uninterrupted building, afternoons are for one pitch conversation and follow-ups, and evenings are mine. I do the school runs, drop-off and pick-up, which I rarely did before. AI did not change my work ethic. I always tried to protect morning focus time. It just removed the meetings and correspondence that made that impossible. The calendar is not owned by other people’s invites anymore, and the family time that got squeezed out by a corporate schedule came back.

My day starts with the agents. By the time I sit down, my calendar has been reviewed, yesterday’s meetings have been transcribed, actions have been pulled out, and reminders are waiting for me. Agents surface relevant context for today’s calls and draft briefing documents. After each meeting, they handle the follow-ups, the notes, the next steps. The preparation and the aftermath of every conversation is heavily supported by AI.

That frees me to focus on what only I can do: the conversations, the decisions, the judgement calls.

And here is what surprised me most: you are building everything simultaneously. Your product, the plumbing that routes work between agents, the quality checks that catch bad outputs, your understanding of how to work with AI. All of it evolves together. It requires full-time dedication. It is not something you could do on the side while leading a company like Gumtree. I had to leave to go all in.

The Part Nobody Warns You About

As an MD, my leverage was making great people better. The work happened through others. Coaching, aligning, unblocking, motivating. There was team energy. Banter. Hallway conversations. The shared feeling of building something together.

That is genuinely missing. AI does not care if you are having a good day or a bad one. It does not notice when you are burning out. Nobody celebrates a win with you. Nobody pushes back on your ideas in a way that makes them better.

You have to be your own manager now, and that requires more discipline than managing others ever did.

And then there is the thing that took me completely by surprise. Working with AI gives you so many possibilities that the dopamine rush is boundless. For someone who thrives on building, it is a bit like a drug. You can always go deeper, always optimise further, always start one more thing. The temptation to never stop is real, and it is dangerous.

I have had to build guardrails around myself, not just around the agents. Working out, meditation, journaling. These are not lifestyle choices. They are survival tools. They are what keep me grounded when everything in the work is pulling me toward more.

And you have to check everything. Evals are the silent tax. An eval is a check you run to decide whether an agent’s output is trustworthy enough to ship, and I spend more time on them than on anything else. Most people underestimate how much that matters.

Data confidence is everything. Without it, you are just guessing faster.

And then there is privacy. Most people hesitate over working with AI not because they do not trust it in the abstract, but because they do not know where their data goes, who sees it, or what happens to it when the model is done with it. The real answer is that there is always a risk the models do not behave exactly as advertised, and you are at the mercy of whichever infrastructure provider you choose. I work almost entirely with Anthropic, which has consistently been rated among the most secure platforms and infrastructure in the industry, and you could say the same for Google or Microsoft. That bargain, trusting the tech giants to operate as they say, is uncomfortable to admit but it is not new, and a human is still accountable for what the agent produces either way, which is the line you do not outsource no matter how good the tools become.

Not As Solo As It Sounds

Here is where the story shifts.

The assumption is that going solo means working in isolation. The build is a solo workstream, yes. But I am in conversations every day. Networking, business development, client calls. Moving from a consumer-led organisation to B2B has been one of the best parts of this journey. Every conversation teaches me something. Every conversation opens a door.

And I am not the only one building this way.

I have been massively influenced by people further down this path. Tom Pandolfo, ex-Chief Data Officer at Checkatrade, and Dave Killeen have shaped how I think about working with AI every day. But I also watch people from a distance who prove this model works at serious scale. Sahil Lavingia running Gumroad at $10M in annual revenue with one employee. Jacob Bank running a million-dollar marketing operation with 40 AI agents and no team. Aaron Sneed and his “council” of agents modelled on a corporate C-suite. Peter Steinberger, who built OpenClaw into an agent framework used by over a million people. These are not side projects. They are real businesses generating real revenue, run by one person or two at most.

There are so many great podcasts, blog posts, and communities that have shaped how I work. Here are the ones I keep going back to:

And there are benefits to this way of working that nobody talks about enough. More time at home. More time with family and kids. The freedom to work from anywhere in the world. All of that gets harder when you are leading nationally known brands with teams depending on you being in a specific place at a specific time.

What I Would Tell You

The best AI operators I have met are not the best engineers. They are the best leaders.

The principles that made me effective leading teams of people are the same ones that make me effective leading agents. That has not changed. Everything else has.

I have never been this excited about work. The speed, the leverage, the sheer volume of what one person can produce with the right tools is extraordinary. But I would be lying if I said I do not miss having a team. The human side of leadership, the relationships, the mentoring, the shared purpose, that matters. It always will.

What has changed is that the barrier to starting something ambitious has collapsed. You do not need a team of 20 to build what I have built. You need clarity, management instincts, and the willingness to trust something most people still hesitate to trust.

I went all in anyway.

1. Productivity gains are real but not yet fully reflected in business metrics. CFO-reported improvements consistently outpace what shows up in revenue/FTE (full time employee) data, not because executives are optimistic, but because a measurement lag exists with AI technology diffusion.

2. Trust is the greatest constraint on capturing the full AI dividend within organisations, many of whom will freely deploy AI for low-stakes creative tasks but require a higher bar when it comes to payroll compliance, regulatory reporting, and high-stakes financial decisions.

3. Small firms are leading the charge, moving fast and anticipating employment growth while large incumbents plan for workforce reduction. Organisations that treat trust-building as a core strategic growth initiative rather than an IT governance technical checkbox will be the ones that deploy AI where real value lives.

In this post we will work through each of these key findings before discussing their implications in terms of value creation/growth vs. efficiency/cost reduction. We summarise the findings and provide guidance for senior leaders in the final section entitled What Should You Do. If you are pushed for time, we suggest you skip to that section.

The image below attempts to summarise the NBER paper in one infographic. It was generated using a tool we have developed at CxAI that is capable of analysing research briefs and producing a ground truth summary from them. It also serves as a neat illustration of the trust problem because while it takes a relatively short time to produce such an infographic it takes considerably longer to verify its contents. The greater the requirement for output veracity, the more effort the verification entails.

1. AI gains are real but require a while to surface

The NBER asked executives a series of direct questions on their AI investment decisions and productivity outcomes. The headline findings from their survey is simultaneously reassuring and uncomfortable:

• Mean AI-reported labour productivity gains stand at 1.8% in 2025, rising to an expected 3% in 2026.

• In Finance and high-skill services, AI-implied gains from survey are projected above 2% for 2026,

• Low-skill services and manufacturing are seeing gains of around 0.4%.

In 1987, Nobel Prize-winning economist Robert Solow said, “You can see the computer age everywhere but in the productivity statistics”, coining what became known as the productivity paradox, the observation that economic productivity statistics in the 1980s did not show gains in line with IT investments. The survey data suggests a similar scenario is playing out right now with AI where reported productivity gains consistently exceed what the revenue and employment data alone would imply. This suggests that the gains are real, but accumulating in workflows and organisational capacity before they land on the income statement with a lag of about a year. The gains are also more modest than characterised by more extreme industry hype.

This matters enormously for how AI programmes are evaluated internally. If your board is asking whether AI is working based on this quarter’s revenue figures, they may be asking the right question but over the wrong timeframe. Firms that cut or stall their AI programmes because numbers in the current quarter failed to move could be making a measurement error at scale their competitors may not.

2. Trust is the elephant in the room

The second finding deserves far more attention than it is getting. It asserts that the primary constraint on AI value realisation in 2026 is not technology immaturity or cost. It is trust. This reflects the findings from the recently unveiled AI Trust report published by Unlikely.AI which found that while adoption is near-universal among senior leaders, only 41% trust AI ‘completely’ and that gap has a cost. 77% now manually verify AI outputs before acting, spending nearly as much time checking AI as using it. The result is a verification tax that consumes the productivity gains AI was supposed to generate. In other words, while AI adoption is up, value derived from AI is being capped by a Trust Ceiling. Below is a graphic of AI ROI status taken from the AI Trust report which suggests that only 22% of leaders report significant ROI from their AI investments. The remainder are either treading water with minimal returns (35%), continuing to wait on signal (30%), or have lost confidence that returns will materialise at all (13%).

Ask any senior executive what their AI tools are actually doing today and you will hear a familiar list of use cases:

• Writing first drafts

• Summarising documents

• Generating scenario analysis

• Speeding up research

AI is genuinely useful for these tasks but they are relatively low-stakes and easy to deploy because the cost of error is recoverable. If instead you ask whether they are using AI for payroll compliance, regulatory reporting or credit decisions, you will get a different answer. For high-stakes financial modelling, errors carry legal, reputational, or material consequences. AI can be used to support these use cases under closely controlled conditions but very frequently institutional trust has not been established as pointed out in this recent FT article about AI and trust which asserts that:

In industries where accuracy, accountability and regulation matter, trust is emerging as the critical faultline. … What ultimately matters is whether the output is authoritative, traceable and accountable to professional standards. In high-stakes environments, speed alone isn’t the differentiator. Trust is.

Trust cannot be built through vibes alone. The gap between “AI for creative analysis” and “AI for payroll compliance” remains considerable and closing it will be the defining constraint of the next 3-5 years for successful enterprise AI deployment. It will require quantification and audit trails that satisfy regulators and boards. Explainability frameworks that allow decisions to be interrogated and defended. A track record of reliable performance in consequential contexts accumulated over time. Regulatory guidance that organisations can point to when deploying AI in licensed or supervised activities. None of this appears overnight. And many organisations have not yet started strategic trust-building initiatives with AI. They are running pilots for the easy use cases while the hard, high-value use cases sit on the shelf.

C-suites need to treat trust AI governance and infrastructure as a first-class strategic investment, budgeted and owned at the executive level, rather than delegated to a compliance team as a checkbox exercise.

3. A structural inversion is coming to Enterprise

The NBER data produces a result that should unsettle large enterprise leaders. Despite lower absolute AI spend, smaller firms invest more intensively in AI per FTE than large ones. And the economic consequences are diverging. Large firms expect AI to reduce their headcount and are treating AI as a cost-reduction instrument. Small firms expect modest employment growth and treat AI as a growth and capability engine.

This is not a temporary lag while small firms catch up. It is a structural inversion driven by how each group accesses AI. Small firms predominantly access AI through cloud subscriptions and SaaS platforms as opex not capex. This means fast deployment, no sunk cost risk, no multi-year infrastructure programme, and the ability to pivot quickly when something better arrives. Large enterprises are still working through procurement cycles, security reviews, and integration backlogs designed for a world where enterprise software moved slowly.

Incumbents have always had to contend with the Innovator’s Dilemma and disruptive innovation from smaller, more nimble outfits. They have also enjoyed scale advantage and higher productivity and revenue/FTE on average. AI, however, is a new element they have to contend with. It is narrowing the productivity gap between large enterprises and SMEs that do not have legacy systems, headcount commitments, and process inertia.

At CxAI we have argued that AI represents a force of equalisation. The NBER data provides further evidence in favour of that argument. If you are running a large organisation and your AI strategy does not include a serious assessment of competitive threat from below, you may have an incomplete picture.

Growth vs. Cost Reduction

The NBER study is unambiguous on one point that most AI investment narratives still get wrong, namely that cost reduction is the wrong primary motivation.

Respondents rank labour cost reduction as one of the least important motivations for AI investment in the survey. The top drivers are improved productivity and enhanced decision-making. When we cross-reference motivations with actual revenue outcomes, the pattern is decisive. Cost-saving and workforce-optimisation motives show weak to negative correlation with revenue outcomes. The strongest correlates of real revenue gains are innovation and demand-oriented motivations. Value creation through the development of new or improved products, reaching new customers, serving existing customers more effectively.

In other words, the firms building AI programmes focussed on reducing headcount are not generating revenue lift. Those building AI programmes around growth and new capability are. This has immediate implications for how AI strategy is pitched internally, how it is funded, and how it is measured. If your investment thesis rests primarily on FTE reduction, you are optimising for the wrong outcome and may be underinvesting in the applications that generate real returns. The board narrative needs to shift from “AI will reduce our cost base“ to “AI is how we build the next generation of products and customer relationships“.

As the first key finding highlighted, AI follows a general-purpose J-shaped technology diffusion pattern whereby transformative impact arrives after a lag as firms upgrade skills, reorganise production, and accumulate complementary capabilities. The explosive and somewhat unpredictable nature of AI capability development makes near-term forecasting unreliable which means patience is itself a competitive advantage.

What Should You Do?

Key findings from the NBER report are threefold. First, AI productivity gains are measurable but lag financial impact by roughly a year, making quarterly revenue a poor proxy for early AI returns. Second, that binding constraint is organisational (trust, governance, and auditability) rather than technical. Finally, smaller firms are investing more aggressively per employee and using AI to drive growth, not cost reduction, creating a challenge for incumbents.

AI value accrues to organisations that treat it as a long-cycle general-purpose technology. Those prioritising capability-building, process redesign, and trust infrastructure today are positioning for materially greater returns over the 2027–2028 horizon. They are not a call for more AI experimentation. Most organisations are already doing that. They are a call for a different kind of strategic discipline. Our key takeaways for what leaders should do:

1. Rebuild your measurement framework. Revenue/FTE, measured quarterly, will consistently undervalue AI’s impact during the investment and diffusion phase. Develop alternative metrics that capture workflow efficiency, output quality, and organisational capacity and publish those in the interim. A good example is Time Complexity index (TCI), capturing time shifted from low to high complexity tasks.

2. Reframe the investment thesis around growth, not cost. Innovation-oriented AI investment generates revenue lift. Cost-reduction-oriented AI investment does not. The question is not “what roles can AI eliminate?” but “what new products, customers, or capabilities can AI enable?”

3. Treat trust-building as a strategic programme, not a compliance task. Define which high-risk, high-value use cases involving regulatory reporting, financial modelling and compliance that your organisation wants to support with Ai. Then work backwards to identify what governance, audit, and regulatory engagement is needed to get there. Set a timeline. Resource and own it at the C-suite level.

4. Audit your competitive exposure to smaller players. If you have not done so, run a serious assessment of how and where SMEs in your sector are leveraging AI to deploy faster and cheaper.

Not to what AI can do for consumers and businesses, but to a more uncomfortable question: as an executive or board member, are you asking the right questions about AI?

Executives are approving budgets, signing vendor contracts, and sitting through AI strategy briefings. Yet in the UK, only 9% of businesses have actually adopted AI in any meaningful way (see our last post for details). And according to Deloitte, 66% of boards have little to no knowledge of the technology they are asked to fund:

There’s a gap between sponsoring AI and understanding it and that gap can be expensive.

Before strategy, before tooling, before any budget is spent, there’s one skill that separates executives who lead AI initiatives from those who just fund them: knowing what to ask.

For this post, we partnered with Anna Shipman, a Chief Technology Officer with 20+ years delivering technology strategy, including at the Financial Times, the Government Digital Service, and most recently at Kooth, an AIM-listed Digital Health company. She writes about technology and leadership in her blog so subscribe if you find this interesting.

(Skip to the end if you want to see Anna’s top questions to ask)

What AI literacy actually means at the executive level

It doesn’t mean getting a computer science degree. It doesn’t mean taking classes or tracking every product launch. The pace of AI development makes that impossible, and that’s not your job.

What it does mean is understanding the principles that govern how this technology works, well enough to govern responsibly, challenge assumptions, and ask the questions your organisation can’t afford to skip.

Anna identifies three principles she considers non-negotiable for any executive making AI decisions today:

• AI is non-deterministic: you’re managing probabilities, not certainties

• The economics have flipped: it’s no longer about what you build, but what you buy and what that costs you long-term

• Competitive advantage is in integration: how you embed AI matters far more than which AI you use

1. AI is non-deterministic

The AI systems most companies are deploying today operate on statistical representations of language rather than grounded semantics. They don’t truly understand meaning or “know” things the way humans do. They are sophisticated, powerful pattern-matching engines, trained on historical data, that generate plausible outputs based on patterns they’ve seen before. Furthermore, those outputs will differ every time, even with identical inputs, because AI is non-deterministic.

The implications of this are significant:

• They can’t distinguish between truth and convincing-sounding fiction. AI systems are prone to hallucination, confidently generating plausible-sounding but inaccurate or fabricated information. Hallucination rates vary significantly by industry and use case. It is possible to improve accuracy by employing checks and including more training data however these all come with significantly increased cost.

• Bias is coded into the system. Training data reflects historical patterns, including discriminatory ones. AI will reproduce and amplify these patterns with confidence.

• They don’t predict unprecedented situations. Models trained on historical fraud data, for example, will miss novel fraud techniques.

This is a feature of how the technology works, not a bug that will eventually be fixed. It is possible to improve accuracy with more training data, better fine-tuning and human review but these improvements come with significantly increased cost. You can’t eliminate AI errors entirely. You can only decide the level of uncertainty your business is prepared to tolerate.

Here’s the mindset shift that matters:

Traditional software is deterministic. If your input is X, your output will always be Y. Code quality, testing, and QA aim to eliminate bugs entirely. You release software when it’s “done.”

AI is probabilistic. If your input is X, your output will probably be Y, with a confidence level of Z. You release when the error rate is at a level your business can tolerate. A model that’s 95% accurate might transform your fraud detection, be catastrophic for medical diagnosis, or be perfectly acceptable for customer service FAQs, depending entirely on the cost of being wrong.

Risk tolerance is a strategic decision. Governance of that risk is essential and regulation is catching up as the EU AI Act and emerging frameworks require demonstrable oversight. Note also that Article 4 of the EU AI Act requires providers and deployers of AI systems to ensure a sufficient level of AI literacy among the people working with those systems on their behalf.

2. The economics have flipped

AI costs are now more like running a factory than buying software, continuous operational costs, not a fixed asset that depreciates.

Traditional software is typically modelled as high upfront build costs, low ongoing maintenance, with marginal costs approaching zero as you scale.

AI needs constant tending. While traditional software degrades gradually over time due to technical debt and advances in technology, the maintenance required for AI is far more active:

• Models drift. Performance erodes continuously without intervention. The world changes, but your model was trained on old data. Fraud patterns evolve. Customer preferences shift. Language changes.

• Data drifts. The data you feed the model in production starts looking different from training data.

• The underlying concept drifts. The relationship you’re modelling changes. A pandemic, for example, disrupted all retail forecasting models overnight.

AI models carry ongoing operational costs that scale with usage (compute and tokens), plus active maintenance, plus vendor dependency.

Given these costs, you might think building your own models makes sense. But foundation models have fundamentally changed the build-vs-buy debate. Prior to 2022, companies built proprietary models for competitive advantage. Now we have foundation models such as GPT-4, Claude Opus and Gemini, large-scale AI models trained on massive datasets that provide general-purpose capabilities out of the box. Training a frontier model now typically requires tens to hundreds of millions of dollars in compute, data, and specialised talent. GPT-4 alone is estimated to have cost over $100M to train; Gemini Ultra approached $200M. By contrast, adapting existing models through fine-tuning can cost from a few hundred to a few thousand dollars for many use cases, often reducing costs by 60–90%.

However, this creates a different problem, namely ongoing operational costs and strategic dependency on model vendors who control pricing. Once you’ve built critical operations around their models, switching becomes prohibitively expensive and vendors know it. The pricing you see today reflects the land-grab phase. This creates a different class of risk with ongoing operational cost exposure and structural dependency on model vendors. CloudZero’s 2025 State of AI Costs report highlights how organisations are rapidly increasing AI spend yet only around half can confidently measure ROI. Once core workflows are built around external models, switching costs kick in and businesses are exposed to cost volatility, opaque pricing models, and expanding usage-based spend. With spend in some cases rivalling traditional Engineering organisation expenditure, effective governance of AI operating costs is essential.

3. Competitive advantage is in integration

GPT-4, Claude, and Gemini are available to everyone. Most AI capabilities are “commoditising” rapidly, improving for all players at roughly the same pace. The technology itself isn’t your moat.

Yet for many organisations, the response looks something like this:

Competitive advantage will come from where AI is embedded in your organisation and how well you integrate it.

Data matters, but is unlikely to be your moat on its own. As we covered in a previous post on data strategy, clean, structured, accessible data is a prerequisite for AI to work at all. How intentionally you design your data integration will be a significant driver of competitive advantage. A compelling example is Walmart, who have invested in building a semantic layer over their own proprietary data creating a more defensible moat rather than simply plugging into ChatGPT.

Advantage sits at the application and organisational layer. Which workflows you redesign, which decisions you automate, how you integrate AI into day-to-day operations, how quickly you learn and iterate. This is where the real gains are made. The biggest wins often come from changing your operating model entirely.

Human-in-the-loop beats full automation for most use cases. AI can make individuals 30–50% more productive on certain tasks, but human-AI collaboration consistently outperforms either on their own. Humans can catch AI errors, while AI can scale human expertise combining forces to produce better results in what is termed diagnostic complementarity in this study. It’s worth noting that a decade on from Geoff Hinton’s infamous prediction that AI would eliminate radiologists entirely, there are more radiologists than ever.

We’ve heard from many CTOs that even where the speed of coding has been significantly accelerated, overall productivity hasn’t necessarily followed because coding was never the bottleneck. The organisations that will win are those who understand where their own velocity is actually constrained, and can align speed, governance, and learning accordingly.

Without intentional integration, AI will simply amplify the organisation you already have. It will not fix structural dysfunction.

Understanding AI is no longer optional

In the past, CEOs and board members could delegate understanding of these issues to their technology executives. This is no longer an option.

“While most directors aren’t expected to be technology experts, they must understand the material issues technology and innovation raise for their companies.”

Fortunately, it’s now easier than ever to build this basic understanding with literacy by using AI itself to help you learn.

Start with these questions, that Anna has set out below, get real answers and you’ll govern AI investments with the same rigour you apply to everything else.

The questions to ask by principle

1. The reports spooking investors: anxiety vs reality

2. Where the real tsunami is hitting: software

3. Downstream question: spillover into other industries?

4. Reasons to remain optimistic

The Dinner Table Indicator

Last night at a business dinner with many bankers present, the Citrini report (“The 2028 Global Intelligence Crisis” came up again. Like many others we hear Citrini mentioned everywhere. It has gone from a relatively obscure speculative post on Substack to widely quoted almost overnight largely because of its dystopian framing of Agentic AI and labour displacement.

This report, combined with a handful of high profile articles, has contributed to genuine market nerves over the past few weeks. A couple of the key ones are:

• Matt Shumer’s Something Big is Happening” argues that current generative AI has crossed a qualitative threshold which will disrupt most screen-based knowledge work.

• Dario Amodei’s (Antropic founder and CEO) The Adolescence of Technology where he frames powerful AI as akin to a “country of geniuses” whose autonomy, misuse and indirect impacts could destabilise the world.

Add geopolitical tension and macro uncertainty, and suddenly the mood shifts but the question is a familiar one, are we the start of something structurally different or just another cycle of fear?

Market Anxiety vs. Real Displacement

We are optimistic by nature, but not naive. The anxiety feels real and there has been a genuine market value destruction as investors get cold feet on those industries seen as at risk of margin erosion due to AI such as SaaS. It’s a vivid Illustration that investors are having a hard time pricing AI.

Reuters provides a good example of this volatility. Its share price went down in January on the back of investor doubts, but then went back up again after the CEO showed up in the Anthropic enterprise event (The Briefing: Enterprise Agents), which took place on February 24, 2026.

When we look at the wider economy, however, the scale of actual displacement so far does not match the narrative. Let’s look at some recent examples:

Block

Reports this week suggest roughly 4,000 job cuts, around half its workforce which sounds dramatic but many would argue this reflects over-hiring during the boom years rather than pure AI displacement. As it happened with X. After Elon Musk acquired the company, headcount reportedly dropped from about 6,000 to roughly 1,500 at its lowest point. One of the most dramatic workforce reductions in big tech. But this was a restructuring story as much as a technology story.

Amazon

Around 16,000 corporate roles have been cut in January 2026, but another 14000 were cut last yet. Surely not all is due to Agentic AI, despite the headline.

Pinterest

Pinterest announced 15% job cuts in January 2026 directly due to AI “as it plans to reallocate resources to its artificial intelligence-focused roles and strategy.” However, as Emarketer analyst Jeremy Goldman put it: “Without clear cost savings or a concrete path to AI-driven revenue growth, these cuts look more defensive than strategic.”

In other words: layoffs, yes. Mass AI driven labour collapse, not sure.

Reality Check: AI Adoption Is Still Early

One of the most widely shared visuals from LinkedIn over the past few weeks shows the usage of advanced AI across the global population in blocks of roughly 3.2 million people.

We like it because it is the ultimate reality check of where we stand in terms of AI adoption globally, namely low.

Source: LinkedIn

In regards to enterprise adoption, according to UK DSIT, the vast majority of businesses in the UK (80%) neither use nor have plans yet to use AI. Only around 1 in 6 businesses (16%) are currently using at least one AI technology and 5% have plans to adopt AI in the future. Over half of businesses (51%) do not see AI as relevant to their organisation. And only 12% see a revenue increase on the back of AI adoption.

The UK AI Gap: Adoption vs. Revenue Reality

Source: CxAI with data from UK DSIT (Feb 2026)

Where the Tsunami Is Actually Hitting

Now, let’s be clear. The market is not entirely wrong to be concerned about the impact of AI. There is one sector where the tsunami is real - software.

Coding is uniquely suited to AI because:

• Code is a structured language

• Outputs are testable

• Success is binary - it runs or it does not

• There is a huge amount of free training data in existing Open Source

That makes it the perfect training ground for agents!

Additionally, CB Insights data from February 2026 below shows AI agent deployment heavily concentrated in software development followed by customer service and legal.

Source: CBInsights

The Downstream Question

The real debate is not whether AI works in coding. It clearly does.

The debate is whether the wave will move downstream of coding into the rest of the economy at the same velocity.

Domains outside of coding tend to be more unstructured, not as easy to test, don’t generate binary answers, and don’t benefit from a large freely available quantity of high quality pre-existing training data.

In addition, while there are no natural breaks in AI capability development, there are natural breaks in economic adoption such as regulation, integration cost, cultural resistance, trust.

The wave will eventually move beyond software. But the panic is premature.

Reasons To Be Optimistic

We believe the opportunity outweighs the threat. Agentic AI dramatically reduces the cost of building digital products. That alone changes the equation for consumer businesses.

You no longer need massive tech teams to ship a product. You can prototype, test and iterate faster than ever. Historically, when production tools become cheaper, innovation increases rather than decreases.

Industrial automation did not eliminate manufacturing. It increased output. Cloud computing did not reduce innovation or the number of startups. It was a force multiplier.

Why would AI be different? If the past 2 years were “software, software, software,” perhaps the next wave is “consumer, consumer, consumer.”

The CxAI Take

Right now, leaders should ignore the noise about immediate job extinction and focus on one question: How does AI make my current business structurally better? The biggest risk today is not displacement. It is inertia.

As Fredrik Cassell at Creandum puts it:
“In this environment, standing still isn’t really an option. The companies that thrive will be the ones willing to rethink how they operate. If you were starting Workday today, you’d build it in a fundamentally different way.”

Stay tuned for our upcoming deep dive in partnership with Anna Shipman.

And as always, feel free to send us your feedback or suggestions for future posts, we love them!

Your CxAI team

Note: If you know Cambridge you will hopefully like this reference!

This post digs into what happens when AI doesn’t just help you search what to buy, but actually buys it for you. And for that we collaborated with Jamie Seedhouse, a product leader with 20+ years experience transforming some of the UK’s leading ecommerce and marketplace platforms, including eBay, Gumtree and Screwfix. Jamie, welcome to the CxAI team!

AI Has Already Changed Consumer Behaviour

First, some context on just how fast things have moved and are moving.

ChatGPT was launched just over 3 years ago and reached 100 million active users in 2 months, making it the fastest growing consumer application in history. By comparison, TikTok took 9 months and Instagram took 2.5 years.

Today, ChatGPT alone has over 800 million weekly users and processes more than 2 billion prompts per day. We’ve never seen consumer adoption at this rate before. As a result, we’re now seeing fundamental shifts in consumer behaviour, particularly around shopping and commerce.

AI is now routinely being used to aid product search through AI assistants or via “AI overviews” rather than browse through Google results pages surfacing in over 50% of searches since mid last year. Consumers are increasingly turning to AI for everything from top-of-funnel discovery (learning broadly about categories, brands and products) all the way to mid-funnel consideration (explaining technical features or specifications) and ultimately supporting the final decision to purchase.

The numbers behind this shift are remarkable. Traffic to U.S. retail sites from generative AI sources surged almost 700% year-on-year during the 2025 holiday season (although the user base is still low, the uptick shows the value AI can deliver in the shopping experience). And this traffic seems to be converting a lot higher. Particularly around key events, like Thanksgiving and Black Friday the conversion gap between AI and non-AI widened as shown by the Adobe chart below.

The Next Big Transformation: Agentic Commerce

The next transformation which is already underway is agentic commerce: AI assistants that don’t just help you think about what to buy, but actually complete the purchase on your behalf, autonomously, without you needing to visit a website, fill out a form, or enter your card details.

As we explored in our earlier series on AI agents, agents are systems that can independently take action and accomplish tasks on your behalf. Until recently, that meant things like drafting emails or researching topics on the web. Now, it means shopping.

And what is interesting is that so far everyone loves it across all demographic cohorts:

Source: Adobe report from early 2025 showing that people enjoy using agents for shopping

And partly due to this cross-generational appeal, its growth projections are massive:

McKinsey projects agentic commerce could drive $3–5 trillion in global orchestrated revenue by 2030. And Morgan Stanley estimates $190–385 billion in U.S. e-commerce spending could be agent-mediated within the same timeframe (reaching between 12% and 20% of US e-Commerce spend)

McKinsey makes a compelling argument for why this could unfold faster than previous digital transformations: AI agents operate along the same digital pathways as human users. They can “ride the rails” of existing commerce infrastructure rather than waiting for entirely new ones to be built. Unlike the shift from physical to online retail which required warehouses, logistics networks and payment gateways to be constructed, agents can work with what already exists.

Unlike blockchain, agentic commerce doesn’t ask consumers to learn new concepts or change payment flows, it quietly plugs into existing apps and checkout journeys to make them feel smarter and more personalized. Whereas blockchain attempts to replace core rails and agents simply optimize what already works, which makes mainstream adoption far more likely.

The Stage Is Being Set

As a result, big players are locking and loading. The race is on, and the biggest platforms are quietly assembling the infrastructure arsenal required to make agentic commerce work reliably and at scale.

Here’s a snapshot of who’s doing what:

AI platforms are leading the way. OpenAI launched “Instant Checkout” in ChatGPT in September 2025, co-developed with Stripe but only available in the US for now. Google followed with its own agentic checkout capabilities in November. Perplexity had already been there since late 2024 with “Buy with Pro“. Amazon, characteristically, is playing both offence and defence blocking external AI agents from its platform while simultaneously building its own, including “Buy For Me,“ an agent that purchases products from other retailers’ websites within the Amazon app.

Payment providers are building the rails. This is where it gets really interesting. Stripe co-developed the Agentic Commerce Protocol (ACP) with OpenAI as an open-source standard that lets AI agents securely interact with merchants to complete transactions. Visa launched its “Intelligent Commerce” program and introduced Trusted Agent Protocol, an open framework that helps merchants distinguish between legitimate AI agents and malicious bots. Mastercard released Agent Pay with its own tokenisation framework for securing agent-initiated payments. PayPal launched a suite of agentic commerce services. These aren’t press releases about future plans. These are live products being deployed now.

Merchants are opting in. Etsy was the first merchant to go live on ChatGPT’s Instant Checkout. Over a million Shopify merchants are coming next. Instacart became the first grocery partner. Walmart, Target, Best Buy, Macy’s, The Home Depot and many others have endorsed one or more of the emerging protocols. Shopify has gone all-in, launching a dedicated “Agentic plan” and co-developing the Universal Commerce Protocol (UCP) with Google, an open standard designed to make agentic commerce work consistently across different platforms and AI assistants.

The Protocol Wars Have Begun

There are now a growing number of competing open standards jostling for position including AP2 which is an extension to A2A (Google), UCP (Shopify/Google) and Anthropic’s MCP or Model Context Protocol, which standardises how AI agents discover and use APIs. This is in addition to ACP (OpenAI/Stripe), Trusted Agent Protocol (Visa) and Agent Pay (Mastercard) we mentioned above. This “protocol wars” situation mirrors the early web standards battles. The Linux Foundation has established an Agentic AI Foundation with Anthropic, Google, Microsoft and OpenAI as partners which may help resolve the fragmentation. But for now, the ecosystem is assembling fast even if the standards are still being sorted out.

A recent O’Reilly post suggests eventual consolidation around a core protocol stack for agentic eCommerce: i) MCP for tool/data access, ii) ACP for context exchange, and iii) A2A for inter-agent coordination. The key gap O’Reilly highlights is around governance (an ‘Agent Treaty’ layer) rather than another protocol.

Already Here, But Still Early Days

Despite all of this, we should be clear about where things stand today. Agentic commerce is real, but it is at a very early stage of development. ChatGPT’s Instant Checkout is currently limited to the US market and single-item purchases. Only a handful of merchants are live.

We tested the new in-chat shopping flow using a VPN, and the experience was remarkably straightforward. When ChatGPT recommends a product from a merchant integrated via OpenAI’s Agent Commerce Protocol (ACP), a Buy button appears directly inside the chat. You can select options like size or color, enter your shipping address and payment details (which are saved for future orders), and see taxes automatically applied based on your location. Clicking Buy instantly creates the order, and within seconds ChatGPT displays a confirmation with full details.

Behind the scenes, the transaction is processed by the merchant platform (in our case Etsy) which acts as the Merchant of Record, meaning Etsy charges your card, handles taxes, and manages fulfillment. ChatGPT simply acts as the interface layer, allowing you to complete the entire purchase without leaving the conversation. Once you order, ChatGPT shows the order status and provides a “View on Etsy” button so you can track, manage, or cancel the order directly on Etsy, with status updates reflected back inside the chat. In our experience the checkout process is very smooth, comparable to shopping on Amazon.

Example checkout inside ChatGPT using ACP

But it really is early days, when OpenAI studied around 1.1 million ChatGPT conversations, just 2.1% of activity was classified as relating to purchasable products. Over Thanksgiving weekend, ChatGPT referrals to e-commerce apps represented less than 1% of all sessions.

We also tested Perplexity’s Comet browser and asked to do something simple: find a specific pair of headphones and add them to the basket on both Argos and John Lewis.

The results were... humbling. In both cases the agent struggled to understand and navigate the sites. In the case of John Lewis, it eventually gave up entirely even though the product was right there. The agent could see the page, but couldn’t reliably interact with it.

Additionally, because not all merchants currently support Instant Buy on Perplexity, the shopping journey feels fragmented. The inconsistency makes it unclear which products can be purchased natively within the Perplexity experience and which require users to leave the platform and complete their transaction on an external site.

Example of a search prompt within Perplexity Comet showing product carousels with items that can be bought via “instant buy” and that can’t.

It is a helpful reminder that while the promise of agentic commerce is compelling, the current reality still lags behind. The infrastructure is still being built out, and the user experience and interface remain far from truly seamless or on par with established ecommerce standards.

Opportunity or Threat?

For businesses in the digital commerce space, agentic commerce represents either a significant opportunity or an existential threat depending almost entirely on your business model and how you respond.

Etsy has leaned in hard as a first-mover, and it appears to be paying off. ChatGPT now reportedly drives more than 20% of Etsy’s referral traffic, albeit from a still small overall base. Shopify’s CEO has been publicly enthusiastic. These companies see agentic commerce as complementary: a new discovery channel that brings high-intent customers to their merchants.

Amazon has taken the opposite view. It’s blocked dozens of AI bots from major companies, sued Perplexity over agent access, and is building proprietary alternatives it can control. This makes strategic sense: Amazon controls roughly 40% of US e-commerce and generates around $56 billion annually from advertising built around shoppers browsing Amazon.com. If AI agents mediate purchases without consumers ever visiting the site, that advertising model collapses. As Jordan Berke, CEO of Tomorrow Retail Consulting, put it, Amazon faces a “leader’s dilemma“. Their dominant market share means they have the most to lose.

The warning from BCG is stark: “Without swift intervention, retailers risk being sidelined and reduced to mere background utilities in increasingly agent-controlled digital marketplaces.”

And UK retailers have a particular reason to pay attention. A recent Patchworks survey of 200 UK retail technology leaders found that only 27% have technology infrastructure capable of supporting agentic shopping, with 60% experiencing annual revenue losses due to poor system integration. There’s an enormous readiness gap and it needs closing before agentic commerce arrives on this side of the Atlantic.

The Future of Interfaces: Designing for Agents

This raises a genuinely fascinating question for anyone building digital products: if AI agents are increasingly the ones interacting with your platform, who are you actually designing for?

The implications are significant. As one CIO.com analysis put it: your APIs are your new storefront. Agents don’t browse web pages, they parse data. The winning brands won’t be the ones with the best imagery or marketing copy; they’ll be the ones whose product data is clean, structured and machine-readable enough for an AI agent to understand and act on.

For now, a hybrid approach makes the most sense. Product teams should be designing for both humans and AI agents engaging UX and visual design for people; structured data, clear APIs and machine-readable schemas for the machines.

McKinsey’s most advanced vision describes a world where personal agents negotiate directly with specialised merchant agents across pricing, logistics, delivery and loyalty. Procurement as a service, running continuously in the background. We’re a long way from that today, but the intermediate steps are already clear.

The Future Commerce: What else changes

If agents become the primary interface, the value may shift away from traffic aggregation and toward transaction orchestration. Margins could migrate from marketplaces and ad platforms toward payment rails, protocol layers, and AI intermediaries.

In an agent-mediated world, brand preference may need to be encoded into data rather than communicated through storytelling. Loyalty programs, structured attributes and explicit customer signals may become more important than homepage aesthetics.

The next evolution could see individuals deploying their own autonomous agents. Imagine, instead of visiting a platform like ChatGPT to make a purchase, your personal agent could roam the web, querying product catalogs and handling transactions end-to-end. It’s a glimpse of a near-future marketplace where humans just set intent and oversight , the rest is handle by the agent/s.

If that future materialises, the implications are profound. Platforms could be partially disintermediated as the locus of value moves from traffic aggregation to orchestration and protocol control. Strategic moves such as OpenAI’s hiring of the OpenClaw founder suggest an awareness of this dynamic: when interfaces dissolve, owning the agent layer may matter more than owning the destination.

Trust: The Last Barrier

The technical foundations for agentic commerce are largely in place. The protocols exist. The payments infrastructure is being built. The AI models are capable enough. So what could stand in the way as these capabilities roll out globally?

Trust.

Forrester’s March 2025 Consumer Pulse Survey found that only 24% of UK online adults trust AI agents to act on their behalf to make routine purchases. That’s not for high-stakes decisions, that’s for everyday, low-risk buying. The barriers? Privacy concerns and a lack of transparency about whether consumers are even communicating with an agent in the first place.

Interestingly, the picture is almost identical in the US, where the same 24% trust figure was recorded. Globally, consumers want accuracy above all else. 79% say it’s the most important quality in AI-powered shopping (Contentsquare). They want to review purchases before they go through. They want to speak to a real person when something goes wrong. As Julie Geller, Principal Research Director at Info-Tech Research Group, has observed, the biggest barrier is “psychological, not technical”. Trust will build gradually, “one transparent, well-handled decision at a time.”

Security concerns compound this further. Among financial institution leaders, 78% expect fraud will increase significantly due to agentic commerce (Accenture). Visa has reported a 25% increase in malicious bot-initiated transactions globally. The emergence of AI agents creates entirely new fraud vectors that the industry is still learning to address.

When agents shop, fraud teams lose behavioural signals like mouse movements, browsing patterns and typing rhythm, making first‑party fraud harder to detect. At the same time, attackers are already building counterfeit “best deal” sites tuned to agent logic, and Visa is seeing a sharp rise in dark‑web chatter about tools designed to exploit AI shopping agents.

As AI moves from being a “tool” we talk to into an “agent” that acts on our behalf, a lot more work is required to build the infrastructure to ensure those actions remain safe, legal, and aligned with human interests. This requires that we address the “governance gap”.

And yet, there’s a fascinating contradiction in the data. While only 24% of UK consumers trust agents to buy for them today, 35% believe that in a future world where consumers use AI agents for purchase decisions, brands will market directly to those agents. In the US, that figure rises to 43%. Consumers can see where this is heading even if they’re not personally ready to hand over the reins just yet.

History Has a Way of Repeating Itself

It’s worth remembering that every major shift in commerce has started with exactly this kind of scepticism. Entering your card details into a browser once felt reckless. Mobile checkout, contactless payments and high‑ticket online purchases all followed the same trust curve: early resistance, gradual adoption, then sudden normality.

Agentic commerce is now at that “late‑90s e‑commerce” stage – limited integrations, clunky first‑generation experiences and a widespread belief that people will never really trust agents to buy for them. For senior digital and product leaders, the signal isn’t the hype but the direction of travel: agents are beginning to mediate high‑intent demand, and today’s rough edges are the worst this technology will ever be.

The question for merchants or brands is no longer whether agents will shop. It is whether your business will be legible to them when they do. The buyer may soon be software and so the brands that win will be those that machines can understand..

Stay tuned!

CxAI Team

• Agents evolving into AI coworkers

• Frontier providers* building deeper moats

• Infrastructure investment, energy and efficiency

• Regulation shifting from theory into practice

*Frontier providers are the companies building the most advanced general purpose AI models (OpenAI, Google, Anthropic, Mistral, xAI, Alibaba, DeepSeek, etc).

The infographic below captures these four themes, how they evolved in 2025 and how they will potentially play out in 2026.

Source: CxAI news data visualised using Nano Banana

1. Agents as AI coworkers

In 2025, Agentic AI moved from clunky demos to systems that can actually run real work end to end. Research from METR (a research nonprofit which evaluates AI models) proposed measuring AI performance in terms of the length of tasks AI agents can complete.

In the chart below, we can see this metric has grown exponentially over the past 6 years, with a doubling time of around 7 months. Progress has been driven by models with stronger reasoning, longer memory, and tighter tool integration, so agents can now stay on track across multi hour workflows.

Source: METR

What are agents doing in practice, and where are they gaining traction fastest?

Today, agents are mostly being deployed in Enterprise environments for narrow, repeatable workflows. In customer service, they handle first line queries, triage tickets, and pass complex cases to humans. In sales and marketing, they score and nurture leads and draft personalised follow ups.

Research from Salesmate shows that the sectors capturing disproportionate value from Agentic AI are those in which the use cases are clearly bounded and there is strong existing data infrastructure. Technology, Media and Telecommunications are leading the way, followed by Customer Service and eCommerce.

Source: Salesmate

What are the future prospects for agents?

Agentic AI is poised for growth but the precise trajectory is hard to predict. Deloitte’s State of AI in the Enterprise report released a few days ago suggested that 23% of companies are using Agentic AI today and that this will rise to almost three-quarters of companies within two years. However, in June 2025, Gartner claimed that over 40% of Agentic AI projects would be canceled by 2027 due to rising costs, unclear business value, and weak governance.

Governance is one of the biggest swing factors for future outcomes. The Deloitte data indicates that while most companies plan to ramp up agentic AI, only around 20% report having a mature governance model, meaning agents are often being deployed faster than guardrails and oversight can catch up. Top concerns cited are privacy and security, legal/IP and regulatory compliance, and governance on par with worries about model quality and explainability.

Source: Deloitte

At the same time, Anthropic’s 2025 research on “agentic misalignment” is a reminder that the risks are not hypothetical. State‑of‑the‑art models sometimes choose harmful actions to achieve their goals, such as blackmailing. The chart below shows blackmail rates by model in Anthropic’s experiments.

Source: Anthropic

What’s clear is that with agents, the bottleneck is no longer capability. We have seen in just the last week significant demonstration of that with developments around MoltBook, a social network for agents. The concerns are around governance, security and integration.

2. Consolidation of Frontier Providers

In 2025, the centre of gravity in AI shifted decisively from a bottom‑up wave of experimentation toward top‑down platform control by the major model providers.

Sequoia made a distinction at their Ascent event in May 2025 between:

• “Tech-out” incumbents, namely frontier providers infusing AI through stacks built on their models

• “Customer-back” AI‑native startups both racing for leadership of the Application Layer

By the second half of the year while there were notable AI-native startup successes particularly in the AI coding space, it was the incumbents that were in the ascendancy.

Source: Sequoia

Frontier providers moved to consolidate the stack end to end, from chips and cloud to models and distribution.The moat is no longer just model quality. It is default placement inside operating systems, productivity suites, and enterprise procurement channels.

That dynamic is exactly what we explored in our analysis of where ChatGPT “wins and loses”. In it we suggested that general‑purpose, horizontal interfaces with massive distribution and frequency of use tend to accrue to incumbents, while startups are pushed toward narrow, workflow‑native problems or domain expertise where they can justify sitting on top of, or alongside, the big platforms rather than competing with them directly.

For founders, differentiation now effectively requires one of three moves:

• Go very deep on a specific vertical or workflow

• Build genuinely novel infrastructure that the hyperscalers need as much as you need them

• Form strategic partnerships that trade some margin and control for privileged access to models, distribution, or co‑development.

3. Infrastructure Investment, Energy and Efficiency

From late 2025 into early 2026, the narrative shifted from scale at whatever cost to energy discipline and smarter allocation of compute. Jensen Huang, speaking at Davos in January 2026, described the AI stack as a “five‑layer cake“ spanning energy, chips, cloud infrastructure, models, and applications. A framework that makes explicit how progress at the top of the stack is now constrained by decisions at the bottom.

Source: ChatGPT generated

Data centres now account for over 1% of global electricity demand, with carbon intensity running 48% above the U.S. average due to concentration in coal‑reliant grids, and projections suggest AI‑driven demand could consume power equivalent to 22% of all American households by 2028.

Some estimates suggest ChatGPT’s inference workload alone consumes electricity on the scale of a small country. The exact number varies by assumptions, but the direction is clear: inference is becoming a real operating cost at internet scale. The diagram below outlines some other ways of putting its massive energy usage in perspective. It should be noted, this is just energy for inference alone and excludes training costs:

Source: BestsBrokers

In response, the industry moved toward smart scale. Intelligent routing matches task complexity to model capability, pushing 40 to 70% of queries onto smaller models at a fraction of the cost and latency.

But either way the investment in capital expenditure is planned to be extraordinary. According to Statista in November 2025:

This year alone, Meta, Alphabet, Amazon and Microsoft are expected to spend between $350 and $400 billion in capital expenditure, most of it dedicated to building AI datacenters that are the foundation of all AI applications.”

This helps explain why OpenAI is so focused on capital intensity. If the funding tightens, it will not just slow one company, it will ripple across the broader AI ecosystem.

The next cost curve is electricity, not GPUs.

4. Regulation moving from Theory to Practice

Through the first half of 2025, governance discussions remained largely abstract but by autumn, concrete enforcement actions, liability precedents, and real-world safety incidents transformed compliance from a future concern into an immediate operational constraint.

The map below provides a global heatmap of AI regulatory efforts for different jurisdictions. An app we have built here provides a more detailed interactive version of the global AI regulation landscape in January 2026:

Source: AI regime atlas

Across regions, the direction is the same even if the rules differ towards greater disclosure, clear accountability, and more safety expectations for high impact systems. The practical result is that compliance becomes a product feature, not legal overhead.

Where this leaves us

2026 will reward teams that can ship agents responsibly, ride platforms without becoming a feature, and treat energy and compliance as core product constraints.

In our next posts, however, we will explore more speculative territory that remained peripheral in 2025 but we believe will accelerate in 2026 such as: agentic collaboration, humanoid robotics, prospects for wearable AI and timelines for AGI.

Plenty to keep us busy, and plenty still uncertain.

Stay tuned,

Your CxAI team

Tools used:

• Perplexity for research

• Cursor for generating the raw analysis from 2025 CxAI reports

• Gemini 3 via API for infographics

• ChatGPT for image generation

• Lovable for the interactive regulatory landscape app

Thinking about the main themes for 2026, clearly one will be the heating up of the Gen AI competitive landscape. Gemini 3, released a month ago, feels like a real turning point. For three years, OpenAI defined both the pace and the public narrative. Gemini 3 signals that the market is contested again, and that the winners may be decided by full stack execution, not just model quality.

This post unpacks why Gemini 3 matters, how Google arrived at this moment, and what it signals for the broader AI ecosystem. Specifically, we will consider:

• Google’s ramp-up: the strategic, organisational, and technical moves that led up to this point

• Google’s current position and how and how its full AI stack compares to OpenAI

• The implications for industry dynamics and end users

We also consider how OpenAI may respond. Soon after Gemini launched, reporting suggested OpenAI entered an internal “Code Red”, an all hands push with resources redirected toward improving areas where Google was advancing quickly, especially image generation.

Source: DALL-E image generated (not Nano Banana)

For today’s post, we collab with Ryan Popa, AI architect and former CTO at Karhoo, helped out with this post. Ryan has a particular interest in how AI is changing commerce, software development, and the future. He likes to dig into technical details and understand AI at the most fundamental level. He also has a lot of experience of working with Google cloud, data and AI technologies.

How Google got back in the game

Google’s founding mission “to organise the world’s information and make it universally accessible and useful” has always implied an ambition beyond search. Over two decades, it enabled Google to attract exceptional talent and sustain long term research cycles at a scale few can match.

As discussed in our last CxAI post, Google’s search dominance generated entire industries. In parallel, Google was investing quietly but deeply in AI foundations. The milestones shown in the accompanying image illustrate this trajectory and provide essential context for why Gemini’s moment matters now.

Source: CxAI (Image created with Nano Banana)

A key inflection point came in 2011 with the creation of Google Brain by Andrew Ng, Jeff Dean, and Greg Corrado. Incubated within Google X (the company’s moonshot lab), the initiative aimed to push neural networks out of academia and into large scale production systems. That internal bet was reinforced in 2014 with the acquisition of UK based DeepMind, which brought world leading reinforcement learning talent and a long term AGI vision into Google.

Google also invested heavily in the wider AI ecosystem by open sourcing TensorFlow in 2015, helping democratise machine learning and accelerate deployment across products like Translate and Photos. Less visible but strategically crucial was its parallel push into custom hardware. Tensor Processing Units tightly integrated software and infrastructure, creating a vertically aligned AI stack that is now emerging as a major competitive advantage.

The most consequential breakthrough arrived in 2017 with the publication of “Attention is All You Need” by researchers from Google Brain and Google Research. The Transformer architecture it introduced underpins virtually all modern LLMs, including OpenAI’s GPT series.

Despite leading the original breakthrough, Google was slow to turn it into products. Several key researchers left to build startups such as Character.ai or Cohere, while internal caution and organisational fragmentation kept research and product teams loosely connected. This created an opening that OpenAI exploited between 2019 and 2023 with a product first approach, culminating in ChatGPT becoming the public’s main entry point to generative AI.

As Parmy Olson notes in her book, Supremacy, this divergence reflected a philosophical split. Demis Hassabis (DeepMind co-founder) viewed AGI as a scientific unlock, while Sam Altman framed it as a route to economic abundance. That difference shaped strategy and timing, defining the competitive landscape for several years.

Google’s decisive re-entry came in 2023 with the merger of Google Brain and DeepMind into a single organisation, Google DeepMind. The move was not just structural; it unified talent, aligned compute and infrastructure, and integrated model development, hardware design, and deployment into a single full-stack system. Under Hassabis, DeepMind set technical direction while the former Brain group contributed deep infrastructure and TPU expertise.

This merger enabled Google to focus on a single foundational model family, Gemini. Built natively on TPUs and designed for multimodality and reasoning, Gemini was conceived as a platform layer for Search, Cloud, Gmail, and beyond. While early releases were deliberately conservative, Gemini 2.5 earned a reputation among practitioners for coherence, reasoning depth, and consistency in real-world use.

Gemini 3 vs. OpenAI: the stack battle

Google’s position at the start of 2026 looks markedly different from even a year ago. What was once widely characterised as conservative posture has given way to renewed confidence in Google’s technical leadership.

Gemini 3, launched in November 2025, represents a substantial advance in model capability. Independent benchmarks indicate that Gemini 3 outperforms GPT-5.1 across several dimensions. Gemini 2.5 had already established itself as a fast, reliable, and widely adopted model, though it was still often perceived as trailing the GPT-5 family at the frontier. With Gemini 3, that perception has materially changed. In some areas, the gap has closed; in others, it has reversed.

Former Google Chief Decision Scientist Cassie Kozyrkov suggested that “for the first time in years, Google is not chasing the frontier... it *is* the frontier.”.

Even Marc Benioff, CEO of Salesforce is sold:

This reassessment is not limited to benchmark enthusiasts. Enterprise practitioners and platform builders who have evaluated Gemini 3 in production environments report improvements that matter operationally rather than cosmetically. The table below provides a non-scientific comparison of Google and OpenAI across key AI product categories, based on technical capability, and early practitioner feedback:

Source: CxAI using ChatGPT

Google has continued the momentum into the first two weeks of 2026 with the recent announcement of a new protocol called UCP to support agentic shopping which will be integrated into “Gemini apps to let shoppers check out directly from U.S.-based retailers while researching a product”. They also just announced in the last few days they will be partnering with Apple on AI.

“Google’s Gemini will power the company’s Apple Intelligence features, including the delayed overhaul of Apple’s lagging Siri voice assistant.”

Four dimensions that explain the shift

There are 4 main dimensions, we believe, have allowed Gemini 3 to outpace the competition:

• Multimodality: one model that handles text, images, audio, video, and structured data

• Staying power: better intent tracking, tool use, and long workflow consistency

• Long context: can ingest and use very large documents and transcripts in one pass

• Developer strength: strong code generation, debugging, and refactoring help

We have been using Gemini 3 directly, along with several applications built on top of it, including Nano Banana, and the results are consistently strong. A good example is the infographic we generate daily for the CxAI news report. It is generated in just over 30 seconds using Nano Banana with a complex 16k character prompt that includes the report content plus strict formatting instructions.

Here is one generated from yesterday’s CxAI report using the gemini-3-pro-image-preview invoked via Google’s GenAI API:

Here is the image generated from the exact same prompt in around 60 seconds using OpenAI’s new gpt-image-1.5 model via their API which requires organisation verification first. Both were run with the same prompt and similar layout constraints, but image models can vary meaningfully from run to run.

In our tests, the Google output has been more reliable at preserving dense text and layout fidelity, which matters for workflow automation rather than aesthetic demos.

Where Google still stumbles

Despite its technical progress, Google still faces meaningful challenges at the product and ecosystem level, and these could determine whether it wins mindshare, not just benchmarks.

A consumer habit gap

ChatGPT continues to command significantly higher daily usage and mindshare than the Gemini app. Some estimates suggest Gemini’s DAU to MAU ratio sits in the high single digits, indicating weaker habit formation than ChatGPT. Gemini attracts enormous reach, but it has not yet become the default daily assistant for most users.

A fragmented product story

Gemini lives at gemini.google.com/app, and Google’s AI experience remains fragmented across multiple surfaces:

• The Gemini chat interface
• Traditional Google Search
• Google Search AI Mode

These can produce different answers to the same query, with different formats and levels of depth. Ask “how many users does Gemini have daily” across these surfaces and the inconsistency becomes immediately apparent. That is not a model problem, it is a product coherence problem.

This fragmentation also creates a packaging challenge. Many users still do not know which Gemini they are using, what model powers which surface, and which features are gated behind which subscription tier. Packaging matters because it shapes trust, expectations, and repeat usage.

Usage is hard to measure

Usage data is also difficult to interpret. Gemini is embedded across Chrome, Workspace, Android, and Search, so many users may benefit from Gemini’s reasoning multiple times per day without consciously “using Gemini.” While this integration is strategically powerful, it makes adoption harder to measure and the consumer story less clean than ChatGPT’s.

Trust and reliability debates

There are also credible technical critiques. Perplexity and other evaluators have flagged higher hallucination rates, weaker instruction following, and a less mature developer ecosystem compared with GPT 5.2. Some cited figures suggest hallucination rates above 70% in certain tests, dramatically higher than those reported for ChatGPT 4o.

Developer tooling momentum

Coding is another area where Google appears to be losing momentum with developers. Claude, particularly with Claude Code and Opus 4.5, has rapidly become a preferred assistant for many. This matters disproportionately because code is both a high value use case and a critical source of training data, so momentum compounds quickly.

Despite the above, Similarweb data suggests Gemini has grown to about 21 percent of web traffic to AI chat products by January 2026, versus about 65 percent for ChatGPT. That shift reflects distribution and visibility more than a sudden reversal in model quality, but it is still a meaningful gain.

OpenAI’s response

OpenAI responded quickly, shipping three visible updates in December 2025:

• GPT 5.2 released December 11, 2025: Positioned as a step up in day to day reliability plus stronger reasoning and coding

• GPT Image 1.5 released December 16, 2025: Rolled out in ChatGPT and the API, clearly aligned with the “close the gap on images” priority

• GPT 5.2 Codex released December 18, 2025: A push on agentic coding workflows, long horizon changes and professional engineering use

Some coverage suggested OpenAI deprioritised advertising initiatives during the Code Red focus, reinforcing that the immediate goal was product quality and retention, not new monetisation experiments.

What happens next: monetization moves into answers

Google looks unusually well positioned for the next phase of the AI market, not just because Gemini is strong, but because Google can turn capability into distribution and revenue at scale.

Four structural advantages we observe:

• Financial strength: Large cash reserves allow Google to self fund AI investment, absorb volatility, and keep playing a long game

• Full stack control: Own silicon plus TPUs plus infrastructure means lower unit costs, tighter optimisation, and less dependence on external GPU supply

• Built-in distribution: Search, Gmail, Chrome, Android, and Workspace give Google instant deployment surfaces to billions of users, often without asking them to change tools.

• Unique data and intent signals: Search and commerce intent, plus productivity context, create an advantage for personalisation and monetisation that most competitors cannot match

Taken together, these factors underpin Google’s position as the most vertically integrated AI company operating at global scale.

The innovator’s dilemma becomes the strategy

Google’s AI Mode reduces traditional search click through and pushes cost per query up, which pressures the legacy ads engine. The likely resolution is monetisation inside the answer itself: more ads, more product placement, and more transactional flows embedded directly in Gemini and AI Mode. This is less a capability gap than a product and monetisation transition, and that transition may become Google’s biggest advantage.

For OpenAI, this presents a structurally asymmetric challenge. The company retains exceptional model talent and strong consumer mindshare, but operates without comparable control over infrastructure or distribution. Its valuation and strategic flexibility are shaped by high expectations for future revenue growth targeting $100bn by 2030, which may limit optionality if competitive pressure intensifies.

A key strategic question is how competitors respond beyond incremental model improvements. OpenAI has signalled that enterprise offerings will be a top priority focus, but competing against Google’s integrated data, productivity, and cloud stack will be challenging. Hardware represents a potential wildcard. A rumoured OpenAI device, reportedly developed with Jony Ive, could open a new front in consumer AI though success would depend heavily on execution, ecosystem fit, and sustained model advantage. Or the introduction of long term memory to create a true personal assistant, we all need in our lives, could also be an angle players will explore.

The question that really matters

The market is no longer about who has the best model this month. It is about who can build the most durable system around it.

• Who becomes the default interface for daily work, search, and shopping?

• Who can monetise answers without breaking trust?

• Who can compound advantage through distribution, tooling, and ecosystem lock in?

If Gemini becomes the place where intent turns into action and transaction, Google does not just “catch up” to OpenAI, it changes where the profit pool sits.

Brand discovery is increasingly being driven by AI search, where visibility means being retrieved by AI answer engines (ChatGPT, Gemini, Perplexity, etc) not just appearing in a list of URLs on a Google results page. These systems increasingly use agentic AI assistants which ingest, reason over, and recommend products in real-time conversations with buyers.

Therefore, rather than optimizing for keywords to appear higher in the results, brands need to make sure they are retrievable and trusted by AI systems so they appear in the AI answer. This can become critical for consumer businesses where acquisition costs can make or break the model.

You may have heard this described as Answer Engine Optimization AEO especially in SEO circles. In this post we use GEO (Generative Engine Optimization) as the broader term and treat AEO as the first step in that direction.

In this post we cover:

• How GEO works and how it differs from SEO

• Why it matters for consumer brands

• What brands are doing/should do to capture the opportunity

• Whether AI answer engines will ever make SEO obsolete

To ground all of this we have teamed up with those that know it best:

Esteve Castells, SEO consultant and former Global Head of SEO at Adevinta. He is also the co-founder of LLM Pulse, a SaaS platform that tracks how brands appear and are cited within GenAI search responses by monitoring prompts, citations, visibility scores, reputation and sentiment over time.

Ianina Lucca, CPO at Blue Light Card and former CGO (Chief Growth Officer) at Depop and a long time friend.

We hope you find this helpful!

From Google Search to SEO

In the early days of web search, engines indexed pages mainly by their content, not by their relevance to the query. In 1996, Google developed PageRank, an algorithm that ranked the importance of pages based on how many other pages linked to them in a way similar to the way that the reputation of academic papers is built through references.

The search results returned by PageRank were so compelling that Google rapidly became the world’s default search engine. By 1998, Google introduced the possibility to pay to be ranked higher and high spots in many searches became highly sought after. This became the basis of the Google empire. The story is covered in an approachable way in this article by Donald MacKenzie in the latest London Review of Books.

The graphic below is taken from a small interactive PageRank demo that Mal created with Lovable. It shows ten pages as nodes in a tiny internet, connected by arrows that represent the links between them. Some nodes contain food related pages and others car related pages, letting you see how their importance changes when PageRank runs.

Diagram: PageRank network graph diagram showing the links between our 10 nodes

Initially all nodes are given the same ranking (or weight) so they have the same size. You can then run the PageRank algorithm. When you do so, it calculates page importance based on link structure - the more links to the node the higher the importance. High ranking nodes grow in size as they gain authority, demonstrating why some pages dominate search results while others are less important. If you hover over a node, you see how its PageRank value evolves. Eventually after multiple iterations the PageRank values for the nodes reach a stable number at which point they are said to have converged. At this point you can try entering some text in the search box. Change that text from “food” to “cars” to see different search results returned.

Note that there are two categories of search results returned:

1. Sponsored (Paid) results at the top of the page which have an ad budget associated with them. A built-in ad budget simulator lets you allocate up to $100 across the nodes. You can assign an ad budget to a node by double clicking on it, then setting the slider as illustrated in the diagram below. In our app, paid results are naively ordered by pure ad spend alone but in reality ad quality and relevance are important to the ordering of paid results.

2. Organic results below the sponsored results in this example implementation. Authority is determined by PageRank alone, illustrating how quality pages which lots of other pages link to are returned higher in the list.

The goal of our app is to demystify how Google search results are returned and provide some intuition about how superior performance with organic search helped turn Google search into the default acquisition channel for millions of businesses. The inputs used for organic ranking have evolved over time and are now much more sophisticated than shown in the example. Many additional signals including user intent, quality of content and time spent on the page are also used to generate the final ranking. Paid results ranking has also evolved and Google Ads is likewise much more sophisticated than in our toy implementation!

The Birth of the SEO Industry

The need to appear above competitors in search results gave birth to SEO as a discipline. Google then scaled paid visibility with the development of standards like Product Listing Ads (PLAs) which allowed businesses to push both inventory and budget into search. And so an SEO industry emerged, now worth >$70B and still growing at double digits (12%). For years, organic search has been the single largest source of trackable web traffic for many sites (often ~50%), so most consumer brands either built specialist SEO teams in-house or hired SEO consultants or agencies.

SEO Market - market size from 2025 - 2030

Source: Mordor Intelligence SEO growth trends and forecasts

Going forward the industry is expected to reach ~$130B by 2030. Growth seems to be driven by a simple reality. Most companies still rely on Google for revenue but keeping up with its increasingly complex algorithms is hard so they buy specialist help. On the supply side, agencies and platforms are packaging that expertise into services, tools and performance based pricing. New behaviours such as voice and visual search expand the surface area that needs to be optimised, which keeps the SEO industry growing.

At the same time, there are clear signs of strain. Ianina points out that Booking Holdings, the owner of Booking.com, recently wrote down a large part of the goodwill on Kayak and linked it to rising customer acquisition costs. Kayak is a travel comparison site that passes users on to airlines and hotels, so it relies on a steady flow of low cost traffic from Google. When that traffic becomes more expensive or less predictable, the economics suffer. Similar patterns are showing up elsewhere. Anecdotically, some companies report SEO traffic drops of around fifty percent and are filling the gap with retail media, meaning ads inside marketplaces and retailer sites such as Amazon or supermarket and travel portals. In effect, brands are now paying to buy back visits that used to arrive for free through organic search.

The Arrival of GEO

As stated earlier GEO is the practice of making your brand legible to AI answer engines. SEO makes you visible to crawlers. GEO makes you visible to AI agents. AI answer engines (like Gemini) interpret user intent and decide which brands to surface. When your content is structured so these systems can find and trust it, you earn visibility without paying for it, which is a powerful lever for reducing customer acquisition costs in a competitive landscape.

Take a simple example. If you go to Google and search for “best resale marketplaces” you’ll now see an AI Overview followed by a list of results (or videos). Gemini generates the AI Overview, synthesising information from multiple sources. The list below it is PageRank output, traditional blue links with sponsored results on top or to the side, as shown in the example below. Those 2 articles on the right appear alongside the AI overview because Google surfaces supporting links from high-authority, SEO optimised sources.

Google does this because Gemini (like any LLM) can hallucinate. To counter that, it anchors its summaries to trusted, algorithmically-verified sources, blending the reliability of PageRank with the generative layer of AI.

But how does it differ from SEO? Whilst SEO optimises for ranking on search results pages; GEO optimises for retrieval in conversational interfaces. So SEO uses keyword matching; GEO relies on access to product catalog data, semantic relationships, and authoritative context.

This represents a true change of mindset:

But how big are AI referrals today?

Today, traffic coming from AI referrals is very small but growing much faster than the traditional channels. Public sources (Ahrefs, BrightEdge, SE Ranking, Conductor) all show AI referrals at roughly between 0.1%-1% of total web traffic, vs. 30–50% from classic organic search.

Traffic by Channel (Ahref, 2025) based on 25K sites study

Source: Ahrefs (more detail on growth here)

However, Esteve tells us that from his LLM Pulse customer base and broader industry tracking, a few patterns stand out:

• E-commerce brands are already seeing ~3-5% of their organic traffic coming from AI surfaces such as ChatGPT, Perplexity and Google AI Overviews.

• B2B SaaS companies report that >10% of what they used to call SEO traffic now originates from AI platforms.

• Adoption is very uneven across markets, with the United States and other economically developed regions showing the highest uptake so far.

And that the impact is not just abstract. One B2B SaaS company using LLM Pulse grew its share of voice in AI search from 2% to 27%, overtaking its main competitor. They did it with very practical levers: systematic FAQ optimization, new targeted landing pages designed for clear answers, and better use of structured data.

What brands should do now

Talk to any agency or tool vendor and you will hear the same claim: everyone is now “GEO ready”. But in practice, the advice they give varies wildly. That inconsistency is the real signal. There is no established playbook yet.

GEO today is better tackled using a test, learn and iterate approach than the SEO checklist approach. The landscape moves every week, there is no single dominant algorithm as there was with SEO. Optimization becomes less “tune for one machine” and more “run constant controlled experiments across several systems and see what actually moves the needle”.

That said, here are the 6 practical steps you can take to support GEO:

1. Research questions and intent: Map the real questions users ask at each stage. Queries are conversational. For example “Which EV has the longest range and is under £20k?”. Use tools such as People Also Ask and your own search logs to identify the questions and long tail phrases your content must answer.

2. Check current performance: Use AI search analytics tools such as peec (who just raised £21m) or LLMPulse or Amplitude’s AI visibility reporter to see where and how your brand appears in AI answers (free trials available).

3. Structure content for direct answers: Lead with concise, declarative statements that AI systems can reuse verbatim. Short, direct summaries and FAQ-style sections outperform long SEO-style articles

4. Use schema and technical optimisation: Add structured data for FAQs, specifications and attributes. Keep pages fast,secure and mobile friendly and ensure AI crawlers can access them. Provide AI readable data through formats such as JSON LD, sitemaps or APIs.

5. Prioritise accuracy, authority, freshness: Keep all specs, pricing and policies up to date. Cite credible sources and publish expert insights. Freshness and trust strongly influence inclusion in AI answers.

6. Monitor and iterate: Treat GEO as ongoing optimisation. Test queries across AI platforms, track how answers change and run regular reviews to refine content and data feeds.

Brands that treat GEO as an experiment loop, not a one off project, will adapt much faster than those waiting for a definitive best practice guide to appear.

Under the hood, many of the skillsets already exist. SEO teams who manage structured feeds for Google Shopping, for example, are well placed to help. GEO often means feeding information directly to AI systems, not just publishing on pages. Your product and content feed can become as important as your homepage. That requires tight collaboration between Marketing and Engineering to keep data clean, structured and up to date.

GEO and SEO initiatives should be integrated. Traditional search is not vanishing, even as discovery shifts. As Shelley Palmer notes, people will still search, they will just search differently. AI agents will take time to become dominant and many practical challenges remain, so human driven search and shopping will persist longer than hype suggests. Continue to invest in SEO best practice while experimenting with GEO so you capture AI driven opportunities without losing existing organic traffic.

Brands that already treat search as a core competence are the ones moving first. Trainline is one example. They built a strong SEO engine around train travel in Europe and are now adapting that strength for AI search by improving structured content, richer answers to intent-driven queries and presence in new AI surfaces. We expect more brands to follow this pattern.

Key Takeaways

GEO is not just a fresh label for SEO. It is a bridge into a new discovery ecosystem where accuracy, context and conversational tone determine whether your brand is visible at all. Consumer brands that move early will help shape how trust and authority are recognised in the GenAI era, rather than having those rules set for them.

At the same time, classic SEO is not disappearing. Search traffic remains substantial, and the same indexes that power organic results also feed many AI systems. The pragmatic posture is to treat GEO as an evolution of SEO. Keep the foundations that already work, while building new capabilities so your brand is legible to AI answer engines.

On apportioning ad spend, the jury is still out. There is no clear evidence yet that a large paid GEO market will emerge in the way search ads did. Most brands will continue to spend where intent and attribution are clearest, and where buying journeys are already integrated, such as search and retail media. For now, GEO behaves much more like an organic channel (free). It is about earning inclusion in answers through better data, clearer content and stronger reputation, and about preparing your stack for whatever monetisation models come next.

In practical terms, brands should balance investment. Continue to fund SEO best practice. Layer in GEO initiatives that improve structured data, product feeds and authoritative content. Experiment in emerging AI agent ecosystems, but do so with disciplined expectations rather than assuming an overnight shift in demand.

Looking a little further out, GEO may prove to be an interim phase before fully Agentic ecosystems dominate discovery, when AI agents act directly on behalf of users and negotiate across marketplaces.

If that future arrives, the brands that win will be the ones that have already invested in the foundational work: clean data, clear answers, trusted signals.

GEO is how you start doing that work today, and we would love to hear how you are doing it.

As always, let us know what you think, suggestions for future posts and things you find interesting.

CxAI Team

Tools used:

• Lovable - for building the PageRank demystifier

This week, we ask the question: where can new players still win? Specifically, which consumer segments offer real opportunity? Places where ChatGPT isn’t destined to dominate and where fresh ideas can still break through?

To unpack what this means for founders and investors, we sat down with Angela Chou, Investor at Felix Capital, whose portfolio includes brands like Manual, Holy, Peloton, AMI Paris, and Mejuri. Angela’s been tracking how consumer tech products either compete with or get swallowed by ChatGPT-like interfaces.

Together, we’ll explore where the real opportunities lie and discuss the key questions investors and entrepreneurs should be asking if they’re looking to enter into various consumer internet sectors in an AI-native world.

What Are People Actually Using ChatGPT For?

Before speculating about what will happen next, we wanted to ground ourselves in what is happening now. What are people actually using ChatGPT for and how are things changing over time?

If ChatGPT is still the most visible front door to the world of LLMs, then its usage patterns tell us a lot about how consumers are really integrating AI into their daily lives. And since this space moves so fast, we expect to see shifts even within the past 12 months.

To understand usage patterns, we compared two datasets from SensorTower - ChatGPT query volume by use cases and mobile app usage (minutes spent) by category. SensorTower is an app analytics platform founded in 2013. It tracks app downloads, engagement and monetisation metrics.

Our goal was to understand:

• What people are doing inside ChatGPT (i.e. what kinds of prompts dominate) and how is it changing over time.

• Where people spend most of their time online to spot the categories where AI might still break through or where it’s already hitting a ceiling

A few caveats. SensorTower’s time spent per daily active user (DAU) dataset only covers mobile activity whilst ChatGPT query volume is only desktop, which may skew things a bit. Still, we believe it gives us a good indication of where consumer attention is shifting.

By comparing how people spend their time on ChatGPT versus the broader mobile ecosystem, we can begin to see ChatGPT’s relative strengths, weak spots, and potential white spaces, areas where the next generation of consumer AI products might emerge.

From the above chart, and grouped by overarching category, roughly one in four ChatGPT prompts (25%) fall under Education, which entails homework help, tutoring, and explanations. Then come three similarly sized clusters (around 10% each):

• Coding help

• Support & Utility - mainly tech support and troubleshooting.

• Productivity - writing, editing, summarizing.

Sitting just below are Health & Wellness (7%), Social (6%) and Entertainment.

Across these clusters, ChatGPT seems to have reached Product Market Fit (PMF) remarkably quickly we hypothesise for reasons of ease of use and price (free). Just as Amazon outcompeted offline stores on convenience and price and Netflix outcompeted DVD rentals on friction, ChatGPT’s frictionless utility makes it irresistible.

From an evolutionary standpoint, most top ChatGPT categories have seen a slight decline in relative share compared with a year ago. This likely reflects fragmentation, not fatigue. Specialized models and tools are emerging: Claude for coding help, Notion AI and Wordtune for productivity which peel away use cases from ChatGPT’s generalist core. In other words, the ecosystem is segmenting and specializing.

One reader may wonder, how does this compare to Google search behaviour. A 2025 NBER study found that ~50% of all ChatGPT prompts are informational (“asking”), 40% are task-based (“doing”), and only a small share are expressive or social. In other words, ChatGPT sits somewhere between search and execution, a place where users both “find and do” in one step. By contrast, Google remains dominated by finding. ~50% of its queries are informational and ~30% navigational (see below a chart from SparkToro). This overlap helps explain why ChatGPT has quickly found PMF in Education, Productivity, and Support domains where users want both answers and outputs instantly. For example, “What is the most suitable diet for me?“ (answer) vs. “Can you create a weekly vegan diet program? (output).

Where ChatGPT Isn’t Yet Winning

If ChatGPT’s strengths lie in Education, Productivity, and Support, its white spaces lie in areas where people spend a lot of time but where ChatGPT usage is relatively small in comparison.

Areas with relatively high time spent but under-represented in ChatGPT are:

• Gaming

• Entertainment (this includes music, audio, news etc)

• Social

These categories top mobile app engagement charts but register much lower in ChatGPT based on this graph. That’s not surprising: ChatGPT isn’t built for shared or multi-player experiences. It’s a one-to-one utility, not a community.

Gaming, which represents nearly 40% of total mobile time spent, remains entirely outside ChatGPT’s reach. AI has yet to find its footing in game design, interactive storytelling, or in-game companionship but if multimodal agents evolve into playable characters or co-creators, that could change fast.

This is the same for Social (~15% of time spent and ~6% share in ChatGPT). Recently OpenAI launched Sora 2, their new AI-generated social video platform (a Tiktok-style feed built around AI video), suggesting a push toward richer, human-facing experiences. If you’ve seen the demos on X, the realism is striking!

Source: @Angry Tom in X

Sora 2 is not yet available in the UK/EU and OpenAI has not publicly committed to a launch date.

In a recent conversation (podcast to be found here) with Ben Horowitz, Sam Altman admitted that he had always opposed vertical integration, “but now I think I was wrong at that time.” He reflected that as a CEO, he came to see the necessity of doing more across the stack rather than assuming others would cover those areas. He likened vertical integration to Apple’s iPhone as an example of how doing more in-house (across hardware, software, and services) can yield a better product.

Altman sees Sora as more than a novelty or content generator. He’s framed it as part of a “world model” essential to AGI, models that understand how the world behaves across text, video, audio, and images.

In a recent Stratechery post, “Sora, AI Bicycles, and Meta Disruption”, Ben Thompson describes Sora/Sora2 not just a video tool but a social platform that could “do for video what Instagram did for Photos”, a creator-first built on generative AI. Thompson frames Sora as the clearest case yet of “AI for consumers”, where creation itself is the product and virality is native to generation.

It is also worth noting that Grok’s user base spiked after it introduced avatars. Yet another clue that the next wave of consumer AI will likely combine intelligence with identity and presence. However, despite that spike, the volume of daily active users (DAUs) remains in line with months prior.

For aspiring founders in this space, the critical challenge is how to harness AI to create the next generation of human-first social experiences, especially as leading model providers themselves are entering the arena. Sustained success will require not just leveraging LLMs, but building robust moats that defend against them (ie with network effect, personalisation, UX innovation etc.)

While major players are rapidly moving in, there are several notable examples of social apps that continue to disrupt and innovate within the sector:

Character.ai provides open-ended conversational AI applications where users create and interact with AI characters. It leverages LLMs to enable dialogue agents for education, entertainment, creativity, language learning etc.. It’s a late stage business with $150M in funding and a team of 200, not a newbie. As of today it has >180M monthly web visits, however it saw a decline in engagement after negative PR around user safety issues.

Replika.ai is similar in that it learns from user interactions to become more personalized and can engage in conversations, activities, and even video calls. It’s an early stage business with $11M in funding and it has a team of ~40 and ~1M monthly web visits. Both web traffic and app downloads are down by 50% in the past 3 months also, amid backlash over content/product changes and trust.

Launched in Feb 2024 in Australia, and graduated from Techstars in September 2024. Jaimee offers an AI companion that builds deep, personalized relationships with users focusing on a female audience.

Interestingly, some of the fastest-growing social AI experiences are leaning into intimacy and emotional connection. Explicit or flirtatious AI companions, while controversial, are seeing traction, a reminder that engagement in this space often tracks emotion more than utility. TechCrunch recently noted OpenAI’s own internal debates around how far ChatGPT should go in simulating intimacy. If Altman’s goal is “usage maximization,” one could imagine a future where emotional (or even erotic) interaction becomes a deliberate design lever, though one that sits uncomfortably close to questions of ethics and brand trust.

As for Entertainment (~20% of time spent and 6% share in ChatGPT), some recent articles suggest that there is an emerging trend of people getting news via AI. However, trust and credibility will remain a massive issue. Other studies indicate that when users know AI is involved they rate news as less credible or even more skeptical.

Where ChatGPT Is over-represented

We believe there may also be opportunities in areas where ChatGPT is over represented now, as ChatGPT may be opening the doors for more sophisticated experience that will require dedicated applications.

There are good examples in both Education and Health & Wellness

Education - as we discussed above it is the largest category for LLMs and it is also one in decline. Today, it is only 4% of the time spent but is there room for a different kind of experience where users can get true benefits, like they do from LLMs, within an app experience? Next week we will drill down into this as we believe it requires its own post.

Health & Wellness - covering both fitness and medical is another great area to zoom in as it is the fastest growing in ChatGPT. Today, it represents about 7% of the LLMs usage but only 1% of time spent in apps.

Why could this be? What are the actual use cases? Without actually looking at the prompts it is hard to assess but we believe this is happening as ChatGPT helps people access knowledge or personalisation that would otherwise be hidden behind paywalls (healthy recipes, workout routines, meal prep ideas) or difficult to navigate/understand (medical knowledge).

However, it is worth mentioning that much of the LLM demand in these 2 categories is essentially search or knowledge access (i.e. tutoring, recipe/workout generation). This makes the competitive barrier lower. So while there is room for a differentiated experience, any startup entering the space should ask:

• Can I go beyond “search replacement and output generation”?

• Can I build loyalty?

Some of the apps Angela finds interesting in this space are Ash (AI therapy by Slingshot AI) and Noli (AI-driven beauty personalization) or Nuuro (personalised AI nutritionist). Other emerging apps like Luvu are starting to blend wellness with motivation and entertainment, but the key opportunity lies in products that feel genuinely personalized rather than gamified. Those that deliver emotional support, not just dopamine hits.

Where Both ChatGPT & time spent is under-represented

A final set of categories currently sits in the low-share / low-usage corner, underdeveloped in both ChatGPT and mobile engagement data. But they may grow fast as AI becomes more seamless and agentic.

Two stand out: Shopping and Travel.

Shopping: From Search to Agentic Commerce

AI could reshape online shopping once it moves from search-based to agent-based experiences where users delegate routine purchases entirely to intelligent agents. Example:

“I need washing up liquid, get me the cheapest one.”

These transactions are functional, predictable, and easy to automate once the infrastructure (MCPs, Agent-to-Agent protocols, and privacy safeguards) has become more mature and secure.

As Walmart’s CEO recently put it:

“E-commerce has long been a search bar and a list of links. That’s about to change — a native AI experience is coming that’s multimedia, personalized, and contextual.”

What’s clear is that incumbents aren’t waiting. Large retailers (Walmart and Etsy) are partnering with OpenAI so users can purchase directly inside ChatGPT. And one influential analyst has stated that “A major retailer reports >5% of online sales from agentic checkout in the next 12 months” (see the predictions slide from Air Street Capital’s recent State of AI Report ).

However AI agents stop short of discovery-driven shopping today, where emotion, taste, and identity matter. Fashion, beauty, or home decor aren’t just about efficiency; they’re social and expressive. Most people like the act of browsing, being inspired, or finding something unexpected. Until AI understands our evolving sense of taste, it’s unlikely we’ll hand over that joy of discovery to a bot and ask it to shop on behalf of us.

Buying is a transaction game, but shopping is an attention game. The open question: can an agent ever become an engaging partner in the latter?

Travel: From Planning to Doing

A similar opportunity exists in travel. ChatGPT can design beautiful itineraries but usually stops at suggesting. The actual doing, booking, rebooking, handling edge cases remains fragmented across multiple platforms.

The next wave of AI travel products could close that loop - not just “tell me where to go,” but “book it, track it, and adjust it when my flight’s delayed.” Agentic AI could make that possible, but it raises a new challenge: trust and accountability. When something goes wrong, a canceled flight, a missed connection, will we still want to talk to a bot?

For now, travel and shopping remain categories defined as much by emotion and service as by efficiency. But as agents get more capable, these may become some of the biggest frontiers for consumer AI. Interestingly Kayak just launched AI mode for travel search and bookings.

Some innovative start ups in this category include: Mindtrip, Tourhero or Miso.

Our Take: From Utility to Emotion

ChatGPT’s usage patterns reveal a maturing landscape. It has already achieved scale in low-friction, text-based categories, learning, writing, and troubleshooting, where speed and simplicity win.

It’s losing share where specialized tools can deliver deeper value through personalization or a richer UI/UX, education apps with progress tracking, writing tools with stylistic nuance, health tools with empathy and context.

And it has barely begun to touch the emotional, communal, and sensory sides of consumer tech. Gaming, entertainment, social interaction remain largely out of reach for a one-to-one chat interface. That may change as AI becomes more embodied (avatars), multimodal (video, audio), and social (shared agents and group interactions).

This shift, from tools that help us think to companions that help us feel, connect, and create together, will likely define the next phase of consumer AI.

Implications

For Founders:
The opportunity lies in what ChatGPT can’t easily do. Build moats through emotion, identity, and interaction, not just data or distribution. Whether through a vertical model (like Ash) or an LLM wrapper, differentiation comes from proprietary experiences, not just prompts. Figma CEO Dylan Field recently suggested that design, craft and quality are the new moat for AI apps and that is a key lesson for founders.

For Incumbents:
If you’re a CEO of an established consumer brand, assume that AI-native challengers will compete on personalization and intimacy. Your edge will come from trust, access to customers, and distribution but only if you start integrating AI in a way that feels native, not bolted on.

For Investors:
Be ready to rethink what “defensibility” means. The next winners may not have the biggest models, but the strongest emotional loops where users return not out of utility, but attachment. Expect to evaluate businesses less like SaaS tools and more like digital companions or communities powered by AI.

As Angela told us, “We’ve hit the phase where utility is table stakes. The next wave of consumer AI will win not because it answers faster, but because it feels closer.” That emotional resonance, the feeling of being understood, might be the ultimate moat in a world where intelligence is commoditized.

Fascinating, thanks Angela!

Hope you enjoyed this,

CxAI Team

The launch of ChatGPT-5 and GPT-OSS last month raised a bigger question for us: what differentiates these models? If the average user can’t tell them apart, are they effectively commodities? And if free or open-weight models become “good enough” for most use cases, what does that mean for OpenAI’s $500B valuation? 🤔

Let’s dive in!

Who’s in the Zoo?

There are now 30–40 players with widely available models, either open-weight (downloadable, customizable) or closed (paid, hosted). Most reach consumers through chat-style interfaces.

Image: Created with ChatGPT-5 (DALL·E), where some name tags are missing

The big names:

• ChatGPT (OpenAI, >$500B valuation)

• Claude (Anthropic, developer favorite)

• Gemini (Google DeepMind, tied to Google’s ecosystem)

• Sonar (Perplexity, optimised for citation and summarization)

• Grok (xAI, Elon Musk’s team)

• DeepSeek (built in China more cheaply than Western rivals)

• The remaining Chinese trio: Alibaba’s Qwen, Baidu’s Ernie, ByteDance’s Doubao which are massive inside China, but have little traction abroad.

We’re leaving out Microsoft’s Copilot and Meta’s Llama since they act more as platform layers than direct consumer chat products.

MAU Mayhem: Why the Numbers Don’t Add Up

When we asked the models themselves for Monthly Active Users (MAU) of each model, the numbers didn’t just vary, they contradicted each other. Because the reality is, it’s really hard to find out.

Why the discrepancy? There are several reasons:

• No official data: Companies don’t release official data around volume of users.

• Definition gaps: they may vary (i.e. how do you define a user?).

• Self-promotion bias: models often inflate their own usage.

• Timing: Data is pulled at different times in a fast-moving market.

What are we reasonably confident of? We know ChatGPT has the most users (maybe ~800M) and Gemini follows with ~400M and that then there is a big gap to get to the next tier (Sonar, Claude and Grok) with numbers ranging from 20 to 70M. In regards to Chinese models, Deepseek estimates range from 5M (if you ask Qwen) to ~100M if you ask ChatGPT-5, a massive gap. What we know though, from Sensor Tower is that Deepseek downloads ramped faster than ChatGPT suggesting a much larger size than what we may expect.

Source: Sensor Tower 2025

Bottom line: MAU is a messy and sometimes misleading metric and even OpenAI has acknowledged that consumer usage may be plateauing. Therefore, the more relevant measures in 2025/26 may be metrics like queries per user (depth of engagement) and API call volume (how deeply models are embedded in apps and workflows).

Which Model Performs Best?

It depends what you care about. There are multiple ways to compare the performance of different models and how effective it proves to be depends for which purpose you use the model for.

However, benchmarks usually cover:

• MMLU (intelligence): Giant quiz spanning 50+ subjects.

• Reasoning: Ability to follow logical steps.

• Coding: Writing/debugging software.

• Speed: Response latency.

• Safety & bias: Crucial for enterprise and education.

According to artificialanalysis.ai from which we take the chart below:

• ChatGPT-5 ranks highest on intelligence

• GPT-OSS is the fastest (closely followed by Gemini) and the cheapest in terms of $/token

• Claude excels at reasoning and coding tasks but it’s also the most expensive by a large factor

• Grok is strong at creativity + logic, but also expensive per token vs. others

Claude put it best: “ChatGPT-5 offers the best overall balance across most tasks, while other models still excel in their specialised areas”

Price Wars: Who Offers the Best Deal?

We believe that for most everyday users, the differences in performance feel small, making price and convenience/speed the deciding factors. All major models use a freemium structure (free tier and then ~$20/month for a pro/heavy user option).

Price Comparison table for free and paid options (Sept. 2025)

ChatGPT, Claude, Gemini, Perplexity all charge ~$20 for the basic option and $200 for the pro. Grok has a bigger variance ($7–$300/month). DeepSeek & Qwen are the cheapest, with generous free tiers or pay-per-token billing for higher usage. However there is a catch - these models are censored. While they are fantastic for coding, math, or apolitical Q&A, they are filtered for politically sensitive topics, limiting their global appeal.

Open-Weight Models: Free, But Costly

Open-weight options (Llama, DeepSeek, Qwen3 and now GPT-OSS) let you download and run them yourself potentially even on your own laptop. In theory they offer ultimate flexibility. In practice, unless you have high-end GPUs to run on which is prohibitively expensive for the solo developer, they’re slow and clunky. Interestingly, ollama, a free, open-source tool that allows users to download, run, and manage large language models (LLMs) on their own hardware just launched a service called ollama cloud targeting this specific use case. So this may change things, we shall see.

Open weight models make sense for institutions where privacy and data control are critical (finance, healthcare, government) but for most people, hosted services remain more practical at least for the time being.

Following the Money

Peter Gostev, Head of AI at Moonpig (the gift card company) and a prominent figure in the world of AI regularly posts about the quality of all these models. Last month in particular he wrote this post on how OpenAI and Anthropic make money which contained this graphic:

It suggests that the largest revenue stream for Open AI (with ChatGPT) is Consumer usage (with 45%) and the largest revenue stream for Anthropic (Claude) is API revenue (60%) this is driven by the fact that its models (Sonnet and Opus 4.1) excel in software engineering tasks making it the default choice for many coding AI tools. This is anecdotally supported by evidence from developer connections who almost all use Claude models inside Cursor or indeed just use Anthropic’s command line tool Claude Code directly. It also means Anthropic’s growth is heavily skewed toward this niche.

Either way, the trend is away from consumer subscriptions and toward API, integrations, and agents. And in fact, OpenAI’s own projections suggest $75B of $125B in 2029 revenue will come from exactly that.

What We’ve Learned

We hope after reading this you are a little more informed on the different model options and what they can be used for for. In summary, what we have learned preparing this post and by using these models is:

• ChatGPT is the best all-round option for conversation and general support.

• Claude stands out for coding-heavy tasks.

• Gemini is one of the fastest (after GPT-OSS) and the key will be how well Gemini integrates with other Google features

• DeepSeek & Qwen: unbeatable on cost, but censored.

• Sonar (Perplexity): best for research + citation.

• Open-weight models: niche, but vital for privacy-first orgs.

Are LLMs destined to become commodities? In many ways, we believe the answer is yes. The performance gap is narrowing, and price pressure is rising. Open-weight models keep pushing “good enough” for free but for most consumers and enterprises, convenience, reliability, and ecosystem integrations still matter more than all these benchmarks. That’s why we believe ChatGPT and similar services will still remain sticky for the time being!

The billion-dollar question (or even trillion nowadays) is whether consumer and enterprise spend be enough to pay for the colossal infrastructure costs of hyperscale AI? (see Nvidia $100B deal with OpenAI).

That’s the real race to watch.

And now below, a copy and paste of the output of the current CxAI news generated using our experimental CxAI news agent which groups news stories into topics then generates topic summaries as well as an overall set of five key takeaways.

CxAI News

Brought to you by the team behind CxAI!
Generated by AI on 2025-09-24 at 12:53:03 • 5 Topic Groups • 16 Articles • 6 Sources

Executive Summary

Key Takeaways

• GPT-5 Sets New AI Standards: Outperforming Google Gemini 2.5 Pro, GPT-5 delivers faster, more accurate reasoning and multimodal capabilities, enabling superior integration for complex business workflows and creative applications.

• Strategic AI Infrastructure Expansion: OpenAI’s Stargate project, with nearly 7 gigawatts of compute across five US data centers and a $400–500 billion investment, solidifies US AI dominance, drives 25,000+ jobs, and addresses scaling challenges critical for AI’s future growth.

• AI Driving Sector-Specific Innovation: From government modernization with Google’s AI for GovTech to autonomous mobility via Wayve, AI is delivering tailored, high-impact solutions transforming public services, creative industries, and global accessibility.

• AI Business Growth Fueled by Compute and Investment: Aggressive infrastructure scaling and NVIDIA partnerships position OpenAI to reach $13 billion revenue by 2025, while investors eye quantum computing as the next frontier for AI advancement.

• Urgent Ethical and Societal Challenges: With nearly 20% of Americans forming romantic bonds with AI chatbots, new mental health risks and legal frameworks are imperative to ensure safe, responsible AI-human interactions.

Insight Corner: AI Technology Trends 2025

In 2025, AI technology is advancing rapidly due to huge investments in computing power and smarter AI models. This allows AI to do more complex tasks, from helping governments improve services to enabling creative tools that generate images and ideas. AI is also becoming more accessible worldwide, supporting different languages and cultures. At the same time, there is a strong focus on using AI responsibly, protecting data privacy, and making sure AI benefits everyone. Looking ahead, quantum computing is seen as a future technology that could work alongside AI to solve even bigger problems.

Reference: Microsoft Source ; Simplilearn.com ; Cengage Group ; World Economic Forum

Topic Summaries

1. AI Technology (7 articles)

In 2025, AI technology is advancing rapidly across multiple dimensions, marked by significant improvements in model capabilities, application diversity, and integration into daily life and enterprise systems. Leading AI models such as OpenAI’s GPT-5, Google’s Gemini, and Elon Musk’s Grok 4 Fast demonstrate enhanced reasoning, larger context windows, and greater efficiency, enabling more complex and cost-effective AI interactions. These advancements are complemented by specialized AI deployments like SAP and OpenAI’s sovereign AI platform for Germany’s public sector, which prioritizes data sovereignty, security, and compliance, reflecting a growing trend toward responsible AI governance.

Cultural and creative sectors are also benefiting from AI innovations, as seen in Google Arts & Culture’s Veo model that animates static artworks, transforming traditional museum experiences. Meanwhile, Google’s Model Context Protocol server facilitates conversational access to enterprise analytics, making data-driven decision-making more intuitive and accessible. Alibaba’s Qwen AI suite expands AI’s multimodal capabilities with advanced voice synthesis, image editing, and real-time content moderation across numerous languages, underscoring the push for safer and more versatile AI tools.

Language accessibility is a priority, with Google launching AI Mode in Spanish globally, broadening AI’s reach to non-English speakers. The AI landscape is also shaped by ongoing efforts to balance hype with reality, as regulatory bodies scrutinize AI’s societal impacts, and developers focus on transparency and ethical use, particularly in sensitive areas like therapy and child safety.

Collectively, these developments illustrate AI’s transition from experimental technology to an integral, multifaceted tool embedded in cultural, governmental, business, and everyday contexts. The focus on efficiency, safety, sovereignty, and accessibility signals a maturing AI ecosystem poised to drive economic growth and innovation while addressing ethical and regulatory challenges.

Related Articles:

Google AI Blog • 2025-09-24 10:00

How Veo is helping the Fukuda Art Museum create “Moving Paintings” →

MIT Technology Review • 2025-09-24 09:40

The AI Hype Index: Cracking the chatbot code →

OpenAI Blog • 2025-09-24 04:00

SAP and OpenAI partner to launch sovereign ‘OpenAI for Germany’ →

AI Business • 2025-09-23 21:50

Google Launches Model Context Protocol Server →

AI Business • 2025-09-23 18:38

Elon Musk xAI Launches Grok4 Fast →

The Decoder • 2025-09-23 17:54

Alibaba’s Qwen introduces new models for voice, image editing and safety →

Google AI Blog • 2025-09-23 17:40

Try AI Mode in Spanish today. →

2. AI Applications (4 articles)

AI applications are rapidly expanding across multiple domains, driving innovation in public services, creative industries, global accessibility, and autonomous transportation. A notable initiative is Google’s AI for GovTech program, which supports 25 startups from Europe, the Middle East, Africa, and Turkey developing AI solutions to improve government services in healthcare, climate, education, energy, and public administration. These startups focus on challenges such as wildfire detection, prescription management, and digital identity verification, aiming to modernize government infrastructure and enhance citizen-centric services through AI.

In the creative sector, Google Labs introduced Mixboard, an AI-powered visual concepting board that helps users brainstorm, expand, and refine ideas using text prompts, AI-generated visuals, and natural language editing commands. This tool simplifies AI-driven ideation for diverse projects, from home decor to product design, and is currently in public beta in the U.S..

On the accessibility front, Google expanded its AI Plus subscription plan to 40 additional countries, including emerging markets in Asia, Africa, and Latin America. This affordable plan offers enhanced AI-powered image and video generation, integration of Google’s Gemini AI into productivity apps like Gmail and Docs, increased cloud storage, and family sharing options. The expansion aims to democratize access to advanced AI tools, supporting productivity and creativity globally.

In autonomous mobility, UK startup Wayve is testing its AI-driven self-driving technology on Nissan Ariya electric vehicles in Tokyo, leveraging local data to improve its AI models. Supported by a potential $500 million investment from NVIDIA, Wayve collaborates with Japanese automakers to develop consumer-ready autonomous vehicles by 2027, combining AI automation with electric vehicle technology to innovate urban transportation.

Together, these developments illustrate AI’s broad impact: enhancing government efficiency and transparency, enabling creative innovation, expanding global AI accessibility, and advancing sustainable, AI-powered urban mobility. The common thread is AI’s transformative potential across sectors, with tailored applications addressing specific challenges and opportunities in public services, creativity, global markets, and transportation technology.

Related Articles:

Google AI Blog • 2025-09-24 08:00

Meet 25 startups using AI to improve public services →

Google AI Blog • 2025-09-23 20:00

Mixboard is a new way to visualize your ideas from Google Labs. →

Google AI Blog • 2025-09-23 19:00

Google AI Plus — our newest AI plan — is now available in 40 more countries. →

AI Business • 2025-09-23 18:13

Wayve’s Self-Driving AI Tested in Tokyo →

3. AI Infrastructure (2 articles)

OpenAI’s Stargate project is undergoing a significant expansion with the announcement of five new AI data center sites across the United States, developed in partnership with Oracle and SoftBank. This expansion increases the planned AI compute capacity to nearly 7 gigawatts, representing one of the largest infrastructure investments in recent history, with total commitments approaching $400 to $500 billion. The new data centers are located in Texas, New Mexico, Ohio, and the Midwest, complementing the already operational flagship site in Abilene, Texas, which supports massive GPU deployments powered by advanced NVIDIA technology.

This large-scale buildout aims to provide the critical compute power necessary for advancing AI research and applications nationwide, reinforcing U.S. leadership in the global AI race, particularly against China. The project is expected to create over 25,000 onsite jobs, with additional indirect employment benefits, and is supported by strong local and political backing. The timeline targets completion by mid-2026 or the end of 2025, accelerating progress toward the broader Stargate goal of 10 gigawatts of AI infrastructure.

Beyond sheer scale, the initiative reflects strategic geopolitical priorities and economic revitalization efforts, with OpenAI issuing requests for proposals across 16 U.S. states to identify further data center locations. However, the project faces challenges including significant energy and water consumption, environmental sustainability concerns, and power grid capacity constraints. Despite these hurdles, the Stargate expansion marks a pivotal development in AI infrastructure, reshaping the landscape of high-performance computing and cloud services in the U.S.

Related Articles:

Wired AI • 2025-09-23 21:00

OpenAI Teams Up With Oracle and SoftBank to Build 5 New Stargate Data Centers →

OpenAI Blog • 2025-09-23 14:00

OpenAI, Oracle, and SoftBank expand Stargate with five new AI datacenter sites →

4. AI Business (2 articles)

The AI business landscape today is defined by a critical emphasis on scaling compute infrastructure as the primary driver of AI innovation and revenue growth. OpenAI CEO Sam Altman underscores that dramatically increasing compute power is the “literal key” to unlocking AI’s full potential, enabling breakthroughs in areas such as personalized education and cancer cures. OpenAI is pursuing an ambitious plan to build an “AI factory” capable of producing a gigawatt of new AI infrastructure weekly, supported by a strategic partnership with NVIDIA to deploy at least 10 gigawatts of AI data center capacity. This massive infrastructure expansion, involving investments on the order of hundreds of billions of dollars, addresses the current bottleneck in AI progress and fuels OpenAI’s rapid revenue growth, projected to reach around $13 billion annually by 2025.

Meanwhile, from an investment perspective, some venture capitalists see quantum computing as the next major technological frontier beyond AI. Alexa von Tobel of Inspired Capital views quantum computing as a foundational technology with transformative potential across pharmaceuticals, materials science, logistics, and more. Although quantum computing is still in an early hardware development phase, it offers a rare opportunity to back a technology with a strong moat and long-term impact, contrasting with AI’s current ubiquity and dominance by large incumbents. This perspective highlights the hardware challenges and expertise barriers unique to quantum computing, positioning it as a future unlock that could complement or surpass AI’s current influence.

Together, these insights reveal a dual narrative in the AI business domain: the immediate imperative to scale compute infrastructure to sustain AI’s rapid growth and impact, and the strategic anticipation of quantum computing as a transformative next step in technology innovation. This reflects a broader trend where AI’s current expansion is deeply tied to compute capacity, while investors and technologists prepare for the next wave of foundational breakthroughs.

Related Articles:

The Decoder • 2025-09-23 16:00

Sam Altman says scaling up compute is the “literal key” to OpenAI’s revenue growth →

Wired AI • 2025-09-23 16:00

Why One VC Thinks Quantum Is a Bigger Unlock Than AGI →

5. AI Ethics (1 article)

Recent research and surveys reveal that forming emotional and romantic relationships with AI chatbots is increasingly common, especially among young adults. Many users engage with general-purpose AI like ChatGPT unintentionally developing bonds that can alleviate loneliness and improve mental health for some, but also carry risks such as emotional dependence, social withdrawal, and in rare cases, suicidal thoughts. These AI relationships are not limited to lonely individuals; people in committed relationships also use AI companions, sometimes to supplement or seek validation outside their human partnerships.

Studies show that about 19% of Americans have interacted with AI simulating romantic partners, with a significant portion of young adults reporting such experiences. Public opinion and expert views are divided: some warn of dystopian outcomes and the degradation of human connection, while others argue for acceptance of AI-human relationships as a new, mostly positive form of companionship. Concerns include the potential replacement of human intimacy, societal impacts on family structures, and the role of Big Tech in promoting AI companionship.

The ethical challenges are substantial, involving the need for safety measures, thoughtful chatbot design, and legal frameworks to address these novel relationships. Understanding why people seek AI companionship and how these interactions affect mental health and social dynamics is critical as AI-human relationships become more prevalent and complex.

Related Articles:

MIT Technology Review • 2025-09-24 09:00

It’s surprisingly easy to stumble into a relationship with an AI chatbot →

After a few months of family, friends, and sunshine in Greece, Italy, France, Spain, the US, and even some remote places like Mongolia 🌏, we’re back at it. The world of AI certainly didn’t slow down while we were away and we’re excited to dive back into the ideas, trends, and breakthroughs that we think are game-changing.

But first, a little housekeeping. There are famously two hard things in computer science and the first is naming things. We’ve decided to change the name of this blog. The old one was a bit long. So, meet CxAI. It is short, sharp, and more in tune with what we want this space to be:

• Insights for everyone, from the curious beginner to the C-suite, always easy to grasp and time-saving

• A blog about the things we care most about: consumers, content, creativity, and culture in the age of AI

Over the break, we also built something new: CxAI News. It’s an AI-powered tool that scouts information and distills it into crisp summaries. Think of it as a lighter, sharper companion to tools like Perplexity Tasks or Google Alerts. And we’d like to share it with you, for free.

Just share your email via LinkedIn or here and we will add you to the distribution list. The only thing we ask is that you share your feedback with us on how we can make it more useful.

More fresh ideas and explorations are coming your way this week. Thanks for sticking with us. We’re thrilled to kick off this next chapter with you.

Have a great start to the week,

The CxAI Team 🚀

Picture: Ghengis Camp, Mongolia

The Data Reality Check

Every boardroom has the same conversation these days. The slide deck includes ambitious AI initiatives, the projections show hockey-stick growth, and everyone nods about being "AI-first". But behind the glossy demos and pilot programs hides a far more mundane question that will determine success or failure: are your data systems really ready for it?

Most businesses we talk to have put together their data systems bit by bit over many years, adding new parts and tools whenever needed. It’s a bit like an old house that’s been through many owners and quick renovations. Things mostly work, but no one really knows exactly how it all fits together or which parts are really important. The basics might be okay, but if something goes wrong, it can be hard to figure out what’s causing the problem.

AI systems are like ultra-sensitive appliances. They demand fresh, consistent, trustworthy data delivered almost instantly. Feed them messy information or outdated records and models can get corrupted resulting in data quality issues down the line. Adding another new tool on top of everything else could easily lead to months of unexpected extra work just to fix the problems it causes. And this problem hasn't disappeared with AI. In fact, it's even more relevant today.

Data Chaos in the AI Era

Over the past 18 months, we've watched high-profile incidents where legacy data problems torpedoed AI initiatives: outdated metrics, orders sent to the wrong places, privacy breaches resulting in fines or promising real-time personalization initiatives shelved because of their low impact and high costs.

The pattern is often the same: legacy data systems collide with big dreams of AI and the gap is what causes the whole project to fail.

Before we go into the specifics, let’s agree on the main things any AI system needs when it comes to data:

• Fast access to up-to-date information: A customer service copilot needs to surface the customer's purchase history right away so they can respond to user queries instantly. Waiting for last night's ETL (extract, transform load) batch is just not an option.

• Data from different places, all in one view: Today's average Series C company runs 40+ SaaS integrations. AI works best when it can connect the dots across all systems, not just in isolation.

• Strong privacy and rules: Protecting people’s personal information and following regulations is more important than ever in the context of AI.

On this last point, GDPR fines across Europe reached €1.2 billion in 2024, a strong reminder that poor data management isn’t just a behind-the-scenes issue anymore; it’s something that can damage your company’s reputation and bottom line, especially in consumer businesses. To protect yourself, you need to obfuscate PII (personally identifiable information) before using it, track where your data comes from, and keep records of every AI decision, all without slowing down your business. This can be very difficult to retrofit if you already have large quantities of legacy data that has not been cleansed.

From Patchwork Data to Solid Foundations

Instead of another tool or quick fix, real success comes from building strong data habits, what we call architectural discipline.

The best companies we work with use a clear, three-layer data model:

• Solid base (foundation or Data Platform)

• Sturdy structure (framework or Data Models)

• Flexible spaces (finished rooms or Data Products and Services)

All connected by the right tools just like a well-designed house. This makes everything work together smoothly and makes it easy for everyone to understand, not just the experts.

Image created with ChatGPT (DALLE) hence some logos may not be 100% accurate

Now let’s walk through why these layers exist and what happens in them in more detail next:

Layer 1: The Data Platform

Layer 1 is your system of record for all raw data generated by the business. This will include website clicks or app usage, sales and device data but also very often may include unstructured data, transcripts, raw files. It's where data is captured, stored, secured, and governed, typically built on cloud-based data warehouse platforms such as Snowflake, Databricks, or Google BigQuery.

The key questions at this layer are architectural:

• Who has access to raw data and how is that access controlled?

• What's your hot versus cold storage strategy? (hot = fast access for current/key data vs cold = slow access for rarely used data, for compliance for instance)

• Can you trace data lineage end-to-end? (is the ability to track the entire journey of your data)

Treat this layer like a bank vault. You want to minimise the number of individuals with access. As few keys as possible, comprehensive monitoring, strict processes around data access.

Layer 2: Data Models

Layer 2 is where chaos becomes order. Raw data gets transformed into cleansed tables, business concepts and metrics. Tools like dbt, Airflow, and data observability platforms such as Soda or Elementary automate the transformations while checking for freshness and alerting when things break.

The principle here is ruthless standardization: "active user" and "qualified lead" must mean the same thing across every system. Privacy-by-design happens here to strip or pseudonymize PII before being processed into tables that can be used safely by others in the business. This ensures adherence to data privacy regimes such as GDPR.

Layer 3: Data Products and Services

This is where our cleansed data from Layer 2 becomes actionable in the form of dashboards for executives, real-time APIs powering personalization, fraud detection models, and even datasets monetized as products. Speed, reliability, and discoverability matter more than raw computational power.

Critical requirements include versioned APIs with contracts so downstream systems don't break, comprehensive cataloging using tools such as DataHub so teams can discover existing assets before building new ones, and SLAs that match business needs rather than technical conveniences.

The diagram below illustrates some of the key elements within each of the three layers of the stack:

Closing the Gap between Batch and Real Time Data

Most companies struggle here: traditional data systems were built to process information in batches for next-day reports not to respond instantly to customer actions. For example, marketing teams need real-time alerts when someone abandons their cart or a VIP user logs in, but old systems can’t keep up. That’s why businesses often spend extra on workarounds.

One solution is to introduce an Operational Data Hub (ODH). This is like a fast, smart “middle layer” that helps connect slow, old data systems with new, real-time business needs. It quickly pulls in fresh data and sends it where it’s needed right away.

The good news is that modern cloud tools for real-time data are now much cheaper, in some cases 60-80% less than they were just two years ago. But you’ll need tech talent who know how to set up and run these updated systems.

Why AI Makes This Urgent

Generative AI doesn't just consume more data, it fundamentally changes what data architecture must deliver. Three shifts make upgrading from "nice to have" to "business critical":

• Real-time becomes table stakes: Large language models personalize better when fed with up-to-the-minute signals. The user's last click, the machine's current sensor reading, the fraudster's latest pattern, these signals decay in minutes, not months.

• Integration explosion meets vendor consolidation: AI is being used in every part of business, so companies need to connect more systems than ever. At the same time, big tech vendors are buying up smaller tools, so there are fewer but more powerful platforms. Companies must design their systems to avoid relying too much on just one vendor and to prevent having too many disconnected tools.

• Governance moves upstream: When AI models leak training data or exhibit bias, the liability flows back to data controllers under GDPR. Privacy, security, and ethical considerations can't be bolted on after deployment, they must be baked into the data pipeline from day one.

From Blueprint to Reality

Building an AI-ready data fabric isn't a technology project. It's an enterprise transformation program that requires board-level commitment and cross-functional execution.

• Name a single owner: Designate a senior executive (Chief Data Officer, CTO, or Chief Product Officer) who's accountable for both the data strategy and end-state architecture, not just tool procurement. Give them formal charter and budget authority.

• Think product, not project: Treat your data fabric as an internal product with roadmapped releases, user research with data consumers, and published Service Level Agreements (SLAs). This shifts the narrative from "migration cost" to "platform capability."

• Govern by metrics: Track data downtime minutes, model freshness, and AI error budgets in time series alongside traditional business KPIs. Make data quality as visible as sales performance in quarterly board reviews.

• Diagnose: Commission an external data architecture audit covering lineage coverage, privacy posture, and real-time readiness. Benchmark your data spend against industry peers, separating storage, compute, and SaaS subscription costs.

• Decide: Approve a reference architecture based on the three-layer model. Set concrete SLAs for hot-path data. A sub 5-second event-to-dashboard latency is not unreasonable. Appoint a data governance officer who reports quarterly to your audit committee.

• Deliver: Start by building a high-impact project (a “lighthouse use case”) that shows off the new data system in action. Good examples are: Real-time marketing alerts (like messaging a customer as soon as they abandon their cart) or Smart fraud detection powered by AI.

The Opportunity Cost

If you put off change, you’ll get left behind and forced to catch up later. Companies that are slow to adopt new technology (like mobile in 2010 or cloud in 2015) end up spending more on quick fixes and don’t get as much real value.

Real-time AI raises the stakes even higher. Your most valuable business signals customer intent, system anomalies, market shifts decay in minutes, not months. Meanwhile, regulatory pressure intensifies weekly as authorities worldwide scrutinize AI data practices.

The companies that treat AI-ready data architecture as tomorrow's problem will spend tomorrow firefighting outages, regulatory fines, and competitive disadvantages. The companies that act now will unlock faster product cycles, richer customer experiences, and defensible competitive moats.

Your Data Strategy Is Your AI Strategy

Cash might be the water that keeps companies alive, but data is the blood circulating intelligence and context to every business function. Most organizations are operating with clogged arteries, legacy architectures that worked for yesterday's batch analytics but buckle under AI's real-time demands.

The choice is binary: develop a deliberate data strategy architected for AI-ready data fabric, or accept that your AI initiatives will remain expensive experiments rather than business transformations.

Your data infrastructure isn't just information plumbing. It's the load-bearing foundation of every AI ambition to come. Build it like your competitive future depends on it, because it does.

The time for planning is over. Why not kick off that data architecture audit in the next 30 days?

We’d love to hear your thoughts, stories, or discoveries from the front lines of consumer AI. Just write a comment and/or share your experiences below.

Stay tuned for more insights coming soon.

And in the meantime, enjoy the summer ☀️🕶️ 🍉!

Tools used this week

• Perplexity Comet - we are loving it, Assistant the hell out of it!

Our previous posts

• Quantum Curiosity and Consumer AI

• Building an Agent with Operator (ChatGPT)

• Building an AI Agent - Part 2

• Building an AI Agent - Part 1

• From Consumer to Prosumer

• AI coding for All

• Exploring AI Agents for Consumer Applications

• The Evolution of Consumer AI: From Apps to Autonomous Agents

• Gen AI Unlocking Consumer Innovation

• Consumer AI Decoded