I'm using today's newsletter to share something I've been working on. It doubles as an opportunity to reinforce a point about Complex Machinery.

The point: Complex Machinery is ostensibly a newsletter about "AI and risk." The former term features prominently in every issue; the latter is often implied. If we're to see "risk" as risk-taking, that activity of looking for ups while trying to avoid the downs, then every "here's where AI went wrong" segment is about risk. Risks that weren't mitigated. Risks that were, in all likelihood, not even acknowledged – because the businesses in question had gone full YOLO.

Today's newsletter should make that connection to risk clearer, and should also shed light into how I parse the news these days.

The project: between recovering from my latest book, writing this newsletter, tending to In Other News, publishing the occasional Radar article, and other such fun, I've managed to carve out some time to document the landscape of risk in AI. What you're reading here today is an extremely condensed version of a larger work-in-progress.

Why "landscape?" Because there's a lot going on, only portions of which are visible or obvious because we so rarely look at the larger picture. (I suppose I could have called it "the AI iceberg" but here we are.) To be clear, this isn't some boil-the-ocean attempt to document every possible element of risk-taking in AI. Instead, I'm exploring the past few years' themes to develop a lens through which better evaluate the changing state of the AI field – the opportunities and dangers alike – which will show where risk is building up in the system. Try to see it as a kind of weather report.

This is more of an analyst report than a narrative read. You're welcome to skim around for sections relevant to your work or interests.

I'll split this initial weather report across two issues. Today I'll cover some key terminology, then walk through the upsides of genAI, and wrap up with something we don't often talk about: the downside risks to AI. The next one will land in a couple of weeks, unless some urgent AI news arrives and I have to write an issue about that.

(What I've described here is what holds true as of this writing – July 2026 – based on the resources I've found. There are likely other matters I haven't accounted for because I didn't know of them, not to mention several I've had to trim out for space constraints.)

Future such reports will be much shorter, and will only cover notable changes.

Terminology

It feels cliche to start with terminology, but I'm using terms that mean a lot of things to a lot of people. It's time to get clear.

The term "AI" already had a loose definition, and with the arrival of LLMs it became a lazy shorthand for the generative variety. Here I'll put ML/AI and genAI under the same umbrella and try to specify "genAI" as needed.

Most of this report is about genAI, though some of what I write has been relevant since the days of machine learning. It just took the popularity of genAI to heighten the awareness and increase the reach by some orders of magnitude.

Likewise, the definition of "risk" relies on the context. Sometimes it refers to the possibility of a negative outcome, like when I refer to risk building in the system. Other times it refers to the negative outcome itself. I also use the investment/trading definition of risk, which is a shorthand for risk-taking – the idea of placing a deliberate bet on a future outcome (and putting one's money at stake to do so).

From there, it follows that risk management isn't just about preventative measures, or constantly thinking about what could go wrong, or sitting still to stay safe. It's an active pursuit of upside gains, seeing where to push harder while making sure you don't self-destruct in the process. (But in a world where everyone else only sees what may turn out well, that side of risk management stands out. And can look negative as a result.)

Taking a wider view, risk management is a matter of asking "what should we do?" followed by "should we do it that way?" and sometimes "is this worth it?" It's a map to your preferred outcome. One that steers clear of the rough parts.

Adding to the confusion is the term "AI risk," which is often another term for "AI safety." That field focuses on preventing AI's downsides (especially in the sub-field of "AI x-risk," or existential risk – the idea that AGI might wipe out humanity). Similarly, there's "AI ethics," a field concerned with bias, misuse, and other matters relevant to AI's impact on people and the environment.

Both AI safety (AI risk) and AI ethics are important elements of the overall risk-in-AI picture. But AI risk-taking covers wider territory, as it also concerns itself with the upsides.

And with that, we can start the journey at AI's upside opportunities.

1 - Upside / opportunities

For all the gloom and doom around genAI, there's still some value. If you know where to look.

Datacenters

The so-called datacenter superscalers are retaining firms to scout for land, arrange financing, and establish influence campaigns to win over local politicians and citizens. This money will only last as long as the datacenter boom itself, but it's possible to ride this cash bump so long as you keep an eye on the horizon. Being service-heavy and inventory-light makes these businesses easy to wind down should the need arise. All in all, this isn't a bad place to be.

Datacenter construction workers face a different story. Construction should be good money for the short run, but it has greater exposure to a datacenter downturn. Between citizen protests and shaky financing deals, workers may be left holding the bag if planned projects stop early or don't launch at all.

There's also some research into datacenter design (to improve power consumption and heat dissipation, for example) but this will only last as long as the mega-sized datacenter trend continues. This is interesting work but on shaky ground.

Supporting services for genAI companies

There's plenty of value in the earlier stages of the modeling pipeline, such as providing training data and data labeling services.

genAI companies are also working with markers, lobbyists, and thinly-veiled "research" groups as part of an industry charm offensive. OpenAI went as far as to acquire a popular podcast so it could control the message have a “constructive conversation about the changes AI creates.” This area should be lightweight and high-margin.

Using the tools

Software development is seeing gains as genAI bots can emit working code. The opportunity here is strong for experienced developers. Software product shops face headwinds as prospects may opt to use a model for DIY. Similarly, cybersecurity pros benefit from using bots to probe systems and code for vulnerabilities. The catch is that cybercriminals use these same tools; genAI may have simply sped up the cybersecurity race without creating strong net gains.

Criminals are winning big with genAI-powered fraud, nonconsensual sexual imagery, and similar material as they now have the power to move at greater speeds and create at greater volume. As far as opportunities go, this one is a strong use case and will probably remain so for some time. (This is flat-out illegal, to be clear; but as it's profitable, it counts as a win from using genAI.) Expect criminals to remain at the forefront of figuring out what this technology does well.

Warehouse bots have already proven their value and they likely get even better. Other workplace automation bots are a mixed bag as their success has varied based on the task in question. (To evaluate opportunities, pay close attention to both the value of automating the task at hand and objective tests of a robot's performance.) Autonomous humanoid robots have a vocal fan base but are both far-off from delivery and of questionable value since non-humanoid, task-specific bots are already here and they require far less R&D. Still, all robots require hefty R&D – this space has a higher barrier to entry and longer-range payoff characteristics.

Despite a spotty track record, companies keep trying customer service chatbots and search summarization. I'm sure there's some money in providing these services right now; but if performance doesn't improve, expect buyers to ditch the chatbots for human customer service reps and keyword search.

Product studios (such as software dev firms) should be able to make some money building genAI projects for companies. The plus side is that companies have the appetite and budget right now. The downside is that this interest is currently tied to genAI excitement and not utility. Expect this work to rise and fall with the hype.

Adjacent spaces

Plenty of profitable opportunities exist in quick cash-grabs, such as AI-washing, snake oil, and overselling the capabilities of genAI-based products. Then you have synthetic influencers and mass-producing clickbait for ad revenue.

There is a small but growing set of openly non-AI products that cater to people who want nothing to do with the technology. It's still early so it's tough to say where this will go, but it has the potential to develop a small and fiercely dedicated fan base today and could grow larger should the genAI field fail.

The idea of a ratings agency for genAI models is well within reach but, as far as I know, doesn't exist. Which is unfortunate. Buyers and investors need someone to move past providers' self-congratulatory model stats and define metrics appropriate for specific business tasks. (I've been thinking about this one, myself. Stay tuned.)

Providers and builders

Interestingly, being an LLM model provider or other major genAI company isn't all it's cracked up to be. While they've drawn quite a bit of investment and customer revenue, all of the circular financing deals, rumors of artificially low prices (subsidizing customer token spend), and financial debt put their sustainability into question. There's also the question of how much of that revenue is profit.

Providers continue to distract themselves by hunting for the Next Hot Thing. Not only is this tougher to do under the pressure they've created for themselves through their marketing campaigns and outsized bragging, but it can frustrate and confuse customers. Adding to the pressure, as I write this both OpenAI and Anthropic are preparing to go public. The road to IPO will bring additional scrutiny on their internal operations and finances, and might put their marketing campaigns in regulators' crosshairs.

The risk/reward tradeoff here stretches to both extremes.

Companies that build genAI products share their providers' exposures, and face a few of their own. The immediate concern is that they are still in the early stages of finding what this technology is good for. There simply aren't a lot of repeatable, easily-adoptable, needle-moving business use cases for genAI, so companies have to sort out use cases specific to their business. (That level of individual attention is beyond the scope of this newsletter but certainly available as a paid service.). Looming in the distance is the threat of genAI providers making material changes to their offerings. Realistic, non-subsidized token pricing tops that list.

Companies that understand how to approach genAI as an experiment should do well here. But as most companies have not taken that approach, I'll file corporate genAI adoption as a high-risk, low-reward space that experiences the occasional jackpot moment.

2 - Downside exposures to AI

It may seem odd to mention risks to AI, since risks from AI make so many headlines. It helps to remember that risks to AI can lead to problems in downstream outputs and other uses. This can impact valid use cases (by rendering them ineffective), worsen the weak and ineffective use cases (as they'll be even less appealing), and create other threats (by making intentional misuse easier).

World and markets

The major players form their own market sector, creating an acute concentration risk: if one falters, share prices for all of them may go down in a heap. This could also cascade into other sharp market movements.

The supply chain for computer hardware holds strong exposure to geopolitical risk since products already face short supply and high demand. Chip manufacturing is top of mind here, as growing tensions between China and Taiwan could impact TSMC; but Russia's invasion of Ukraine, plus the current situation in the Strait of Hormuz, can trigger knock-on effects of their own.

Physical infrastructure

Threats to datacenter construction are manifold. Hopeful superscalers face shortages of land, power, and support from local communities. Europe's current heatwave is a reminder that every facility faces climate risk – past a certain temperature, there's no way to cool the machines. Datacenter sites may also fall victim to vandalism or theft, the latter of which can impact the supply chain at the terminus.

These massive outposts also rely on shaky financing that includes circular deals and debt instruments like bonds and loans. Lastly, required hardware (GPUs, disk, memory) is in short supply and prices are climbing. There's still the open question of whether all of these datacenters are even necessary, given that genAI excitement doesn't match its utility. While a market correction may impact borrowers/bondholders and the genAI legend, it might not impact the technology usage on a practical level.

Technical infrastructure

Failure to manage the technical environment poses a risk to people and processes that rely on the models' outputs. Companies need to protect the entire data supply chain against data tampering and model storage tampering, and also get a handle on operational matters like model deployments. Failure to do so could lead to poisoned models and/or useful models going offline. It's hard to tell how large of an exposure this is, simply because companies rarely publish this kind of information. That said, I have a hunch that this is an under-appreciated risk exposure. A failure could lead to widespread damage.

Providers also need to consider model misuse such as unauthorized use or malicious prompts. I'm flagging this as a sizable downside exposure because even the larger, established genAI providers are still having trouble with it.

Skills and planning deficits

A hiring shortage could impact companies that build genAI systems and also those that build on said systems. Larger firms are able to lure hires with outsized pay packages (including Meta allegedly offering $100M to members of its "superintelligence" unit); but competition at those higher levels leaves smaller firms at a disadvantage for hiring. This represents a weird exposure because the talent wars may ultimately be artificial, with employees suddenly losing their job when the market cools.

There's also a skills shortage in companies that use or build on AI, which is compounded by a lack of AI strategy at the executive level. The combined effect is that these companies are unlikely to see a real return on their genAI investment. Most of the risks here impact the companies themselves, though their collective missteps may sour industrywide distaste for genAI – even for valid use cases.

Provider/platform risks

Most AI-backed applications are built on top of models from the major providers, which creates a platform risk for buyers: changes in prices or in backing models can wreak havoc on downstream products and processes. Over time, once the competitive dust settles, it wouldn't surprise me to see collusion or other cartel-like activity among the providers.

The providers pose a subtler risk in that their quest for revenue may drive them to put increasingly questionable and/or ill-planned products to market, with the added risk that they will neglect models and services that customers have come to rely on.

One interesting provider risk stems from the troubles in the leadership team. Any executive-level shakeup can lead customers and investors to question the company's future prospects. Similarly, since the CEO often doubles as the public face of the company, their personal or professional mishaps could put the operation in doubt.

Training data

An inability to acquire training data – the initial dataset or additional pulls – will stymie any modeling effort. LLMs rely on a much larger volume of training data than their domain-specialized siblings. This leads to questionable data collection tactics as companies turn into upstart data brokers, selling anything they can find to LLM providers and other model-builders. (Should the word get out on a dirty data deal, LLM providers may face reputation damage due to the unwanted press coverage.) Thanks to widespread genAI use, it's pretty much inevitable that "slop" and other synthetic materials will wind up in training datasets for future models – leading to an odd case of genAI eventually poisoning itself through model collapse.

Training data also poses psychological harms to human reviewers and data labelers, as workers review the toxic content that the rest of us don't want to see. I expect the pool of data labelers will shrink and/or news coverage of these working conditions will drive the AI companies – who are the ultimate buyers of these services – to change tactics.

Workplace

Companies' over-eager genAI adoption drives a number of downside exposures, notably malicious obedience (employees wasting money and effort on genAI projects that they know add zero value beyond an executives' ego boost) and damaging compliance (employees who mean well but will still cause genAI-related mishaps). Both groups contribute to a growing corporate concern of AI-related cost overruns that stem from runaway LLM token usage and expensive, high-profile project failures.

Liability and other legal issues

I've already mentioned some potential problems in collecting training data. There are also potential legal risks around model usage.

A model's lack of transparency and interpretability can limit its use in regulated areas, such as consumer lending. While this concern isn't unique to genAI, there is the question of whether the LLM provider or the end-customer is on the hook should a regulator come knocking. This legal uncertainty translates to a sizable risk exposure.

The desire to go public creates an interesting set of exposures for genAI companies: the road to IPO opens them to greater scrutiny, especially their marketing practices and their finances. It's possible that questionable and/or unimpressive numbers will raise questions, and pre-IPO marketing rules may limit the companies' ability to smooth that over. And should a company fail to IPO, or should its public debut not lead to a rally in share prices, they could suffer reputation damage.

Changes in mood

Right now genAI faces reduced interest (loss of halo effect) from individuals and businesses, growing anti-AI sentiment from creatives and gamers, the population's frustration over increasingly invasive products (and the companies that release them), sticker shock from buyers as they face daunting bills for their indulgence, and a general promise fatigue due to repeated failure to deliver. Not only can all of this reduce buyers' appetite for genAI, but ML/AI may suffer due to guilt by association.

Collectively, the aforementioned concerns drive the ultimate risk to genAI: when the music stops, will genAI provide enough utility to replace the hype?

More to come

That's it for the first AI weather report. The next newsletter will cover downside risks from AI, as well as a summary of the findings.

In other news …

For more links to recent news, and with a slightly broader scope, I encourage you to check out my other newsletter. It's a weekly, curated drop of what I've been reading.