What's Going On in the World of AI: An Unbiased View

Before anything else, a disclosure, because a piece calling itself unbiased that hides its own position is just a better-dressed sales pitch. We build with these tools every day. We have commercial relationships with the platforms this article discusses, and we make our living engineering AI into businesses, which means we are not neutral about whether AI is useful. Read what follows knowing that. What we do not have is a stake in any particular company winning, and what follows is our honest read of the record.

Here is the difficulty. The AI story reaches you in two versions. In the first, machines are about to reorganize civilisation and anyone not moving is already behind. In the second, it is the largest capital misallocation in living memory and the crash is a matter of when. Both versions are told with total confidence, both are told loudly, and both are told by people who own something. The version you rarely get is the one that separates what is actually proven from what is merely claimed, and admits which parts nobody knows.

So that is what this is. What happened, sourced. What the numbers really say. And where the truthful answer is a shrug.

The scoreboard, and why nobody has run away with it

Start with the thing everyone wants ranked, because the ranking is more interesting than the winner.

The frontier moved four times in five weeks this summer. Anthropic released Claude Fable 5 and Mythos 5 on 9 Jun 2026, then Sonnet 5 on 30 Jun. Elon Musk’s lab, now folded into SpaceX, put out Grok 4.5 on 8 Jul. OpenAI shipped the GPT-5.6 family on 9 Jul. Google announced Gemini 3.5 Flash and a new any-input-to-any-output model at its developer conference in May.

Now the detail worth more than all the launch coverage put together. In OpenAI’s own published benchmark tables for GPT-5.6, on its own website, on launch day, it reports that Anthropic’s Fable 5 scores higher than its flagship on the Artificial Analysis intelligence index, and that Anthropic’s Mythos 5 beats it substantially on a leading software engineering benchmark, 80.3% against its own 64.6%. The market leader published the numbers where it loses. That is a more honest signal about the state of the race than any analyst take, and it tells you the thing worth knowing: there is no runaway. The lead trades hands every few weeks, on different measures, and whoever shipped most recently looks best until somebody else ships.

The rest of the field has sorted itself out in ways that would have surprised people a year ago. Meta gave up on open weights at the frontier, releasing its proprietary Muse Spark in April and effectively retiring the Llama line it built its reputation on. The open-weight banner passed to the Chinese labs, who now carry it: Alibaba’s Qwen under an Apache licence, Z.ai’s GLM under an MIT licence, plus Moonshot’s Kimi and MiniMax. Their benchmark claims are mostly self-reported and should be treated accordingly. Their prices are not a claim. They are a fact, and we will come back to them, because for a business owner the prices are the story.

The money, where both sides of the argument are right

This is where the bubble question lives, and the honest answer is that both cases rest on real numbers.

The spending is genuinely extraordinary. Alphabet guided to $180–190B of capital expenditure in 2026. Amazon is near $200B, Microsoft near $190B, and Meta at $125–145B. Add it up and the big platforms are spending somewhere in the region of $650B to $725B this year on AI infrastructure, roughly double last year. OpenAI, meanwhile, has committed to something on the order of a trillion dollars of compute across a decade with seven different vendors, while its own forecasts reportedly project a loss of about $14B for 2026. Those two facts sit in the same company, in the same year.

And the revenue is also genuinely real. Nvidia’s most recent quarter put data centre revenue at $75.2B, up 92% year over year. Google Cloud grew 63% with a backlog of committed future business above $460B. Microsoft’s AI business is running at a $37B annualised rate, more than double a year ago. Anthropic, in a primary statement accompanying its funding round, reported run-rate revenue crossing $47B in May 2026, having raised $65B at a $965B valuation. Both companies have now filed confidentially to go public.

So which is it. The bubble case says the gap between what is being spent and what is being earned is enormous and widening, that the leases are being structured to keep the true number off the balance sheet, and that the hyperscalers are now borrowing to fund the build. The not-a-bubble case says you cannot call revenue growing at 63% and 92% a mirage, and that infrastructure spending ahead of demand is what every real technology build-out has looked like from the inside.

What actually happened is the most useful data point of all. In late June 2026 the market voted, hard. The Nasdaq fell more than 4% in a session, semiconductor stocks shed over a trillion dollars of value, and Korea’s KOSPI dropped 10% and tripped a circuit breaker. And then, having done all that, the index finished the period still up around 10% for the year. That is not a crash and it is not a coronation. It is the sound of investors changing the question from how much are you spending to what did it return, which, incidentally, is the only question we have ever thought was worth asking.

The story nobody is telling you: it got cheap

Here is the development with the most direct bearing on whether AI is worth anything to your business, and it has been almost entirely drowned out by the drama.

The price of intelligence has collapsed.

OpenAI’s cheapest GPT-5.6 tier launched at $1 per million input tokens and $6 per million output. Anthropic launched Sonnet 5 at introductory pricing of $2 and $10, against $3 and $15 thereafter. DeepSeek made a steep discount permanent and now sits around $0.44 per million input tokens. Xiaomi cut its API prices by up to 99% in May. The Chinese labs are pricing at a fraction of the American ones, and the American labs are cutting hard in response.

Understand what that does. In the arithmetic that decides whether an AI project pays for itself, the cost to run the system is one of the terms. When that term falls by an order of magnitude, jobs that did not pencil out eighteen months ago start to. The frontier getting smarter is interesting. The floor getting cheap is what actually changes what you can afford to automate, and it is happening quietly while everyone argues about valuations.

The week the rules changed

If you take one thing from 2026, take this, because it is a fact about the supply chain your business would be building on.

On 12 Jun 2026, three days after Anthropic released Fable 5, the US government issued an export-control directive suspending access to the model by any foreign national, anywhere, including Anthropic’s own employees. The practical effect, by Anthropic’s own account, was that it turned both models off worldwide, for every customer. They stayed off for nineteen days, and came back on 1 Jul after the controls were lifted.

Set aside who was right. Anthropic’s public position was that the underlying concern rested on a narrow jailbreak whose capability was widely available in other models, and that applying the standard consistently would halt frontier releases across the industry. The government evidently disagreed at the time and then changed its mind nineteen days later. What matters for you is the demonstrated fact: a commercially deployed frontier model, one that businesses had built on, was pulled offline globally by a government, with days of notice, and the vendor could do nothing about it.

This is not an argument against building with AI. It is an argument against hard-wiring your business to exactly one model with no way to switch. The industry has since begun building shared machinery to prevent a repeat, including a joint severity framework for this class of risk developed by Anthropic with Amazon, Microsoft and Google. In the same period, a US executive order established a voluntary pre-release government review for advanced models, and GPT-5.6 became the first frontier model publicly cleared through it. Governments are now inside the release process. That is new, and it is not going away.

The bottleneck was never the chips

The constraint everybody expected was GPUs. The constraint that actually arrived was memory and power.

DRAM prices rose 80% to 90% in a single quarter, with data centres now consuming the majority of global memory output and the manufacturers shifting production to the high-bandwidth memory the AI accelerators need. You can read the consequence directly in the corporate filings: when Anthropic closed its funding round, its new strategic investors included Micron, Samsung and SK hynix. A model lab took investment from memory manufacturers. That tells you what is actually scarce.

Power is the other wall. Grid interconnection queues, transformer lead times measured in years, and rising local opposition are now a real constraint on how fast any of this can be built, and the Stargate project, the largest announced build-out, has reportedly run into partner disputes and delays while OpenAI routes around it with bilateral deals. None of this stops AI. All of it slows the schedule that the valuations assume.

Where the honest answer is “nobody knows”

A piece claiming to be unbiased has to be willing to say this part.

The ROI evidence is weaker than everyone pretends, on both sides. You will have seen the statistic that 95% of corporate AI pilots produce no measurable financial impact. We have cited it ourselves. It comes from a 2025 MIT working paper built on 153 survey responses and a few dozen interviews, and it has been methodologically challenged since. It is now quoted everywhere as though it were a fact about 2026, which it is not. The directional finding, that most pilots fail on integration rather than on model quality, is echoed by enough other work that we still believe it. The precise number should be handled with tongs, including when we are the ones holding it.

The jobs picture is genuinely unclear. Companies attributed 87,714 job cuts to AI in the first five months of 2026, more than in all of 2025, and AI is now the leading stated reason. But stated is doing a great deal of work in that sentence, and the most interesting analysis of it argues that companies are cutting for AI’s potential rather than its measured performance, which is to say that “AI” has become a respectable word for a decision made for other reasons. Both things are probably true at once. Anyone who tells you the split with confidence is guessing.

And nobody knows whether the spending pays off. That is not a dodge. It is the actual state of the evidence.

What a Canadian business is actually governed by

Cutting through the regulatory noise, because most of what you read about AI law does not apply to you.

Canada has no AI statute. The Artificial Intelligence and Data Act died when Bill C-27 lapsed on prorogation in January 2025. The government’s new privacy bill, C-36, tabled on 15 Jun 2026, deliberately leaves AI out and covers privacy only. So the rules that bind you today are the ones that already bound you: PIPEDA, its provincial equivalents, and in British Columbia the public-sector data rules that catch anyone working with a health authority or public body.

The EU AI Act is real and its enforcement powers arrived on 2 Aug 2026, with fines reaching 3% of global turnover, though the obligations for high-risk uses have been deferred to December 2027. It matters to you only if you serve EU customers. If you do, it matters a lot.

The United States is in an argument with itself. A federal executive order is attempting to challenge state AI laws, but it has no preemptive force on its own, so state rules stand until courts say otherwise. If you sell into the US, you are complying with a patchwork.

The short version for an operator on Vancouver Island: nothing in AI law currently obliges you to do anything you were not already obliged to do. That will change. It has not changed yet.

What we think this all means, and what we do not

We build on these tools, so take this for what it is worth, and here is the honest read.

The technology is real and the returns are real in specific places, and both of those statements survive every deflating fact in this article. The valuations may well be too high and the infrastructure may well be overbuilt, and neither of those things would make the useful applications less useful. A crash in AI stocks would be a catastrophe for people holding AI stocks. It would not make a document-intake system that saves a clinic twenty hours a week stop saving twenty hours a week. Those are separate questions and they get deliberately conflated by people who need you to be either frightened or excited.

What we take from 2026 is narrower and more practical. Do not marry a single model, because a government can turn one off for nineteen days and did. Do not buy on benchmarks, because the leader publishes numbers where it loses and the ranking changes monthly. Do watch the price collapse, because it quietly moves the line on what is worth building. And keep asking the only question that ever mattered, which is whether the specific job you are pointing this at actually pays, because none of the news above changes that arithmetic by a cent.

The state of play, in one paragraph

Five or six labs are trading a lead that nobody holds for long. The money being spent is staggering, the revenue being earned is real, and whether the first justifies the second is a genuinely open question that no honest person can settle for you today. The models got dramatically cheaper, which almost nobody is talking about and which matters more to your business than everything else on this page. A government demonstrated it can switch a frontier model off worldwide. Memory and electricity turned out to be the real constraints. Canada has not written its rules yet. And the number everyone quotes about AI failure is from a survey of 153 people.

That is the whole picture as of 23 Jul 2026, as best we can establish it, and the most useful thing about it is how much of it is still unsettled. Anyone selling you certainty is selling you something. What you can actually do, today, is ignore the scoreboard, watch the prices, keep your options open, and judge every proposal by whether the job it is pointed at is worth doing. That worked before this year’s news and it will work after it.