With contributions from Vivien Lee, Aaron Ma and Eric Savoie
The next AI iteration: consumer agents
The artificial intelligence produced by large language models has thus far been primarily used for two things:
To program computers more effectively, with many (mostly) business-level applications that we believe will continue to raise economy-wide productivity growth and corporate profits.
Of greatest direct relevance to individuals, AI has so far amounted to a highly educated chatbot, capable of interacting with the user and summarizing the state of knowledge on a subject. Measured in terms of their sheer intelligence, AI models have advanced by leaps and bounds over the past several years (see next chart and note the log-scale on the y-axis).
Figure 1. AI models are becoming exponentially more capable
As at 05/08/2026. Chart displayed in natural log scale. Estimates above 16 hours are unreliable with METR (Model Evaluation and Threat Research) current task suite. Sources: METR, RBC GAM
This is impressive, but limited. Put simply, the AI brain improved, but it still lacked arms and legs – the ability to interact with the world and actually do things for the lay person.
Happily and remarkably, the latest AI tools – consumer agents such as Muse or Instinct, which were released in recent weeks – solve the arms and legs problem. Supplied with your passwords and credit card number, they venture out into the world, acting on your behalf (see next table).
Figure 2. The next iteration of AI interacts directly with the world
As at 10/05/2026. Source: RBC GAM
Is this revolutionary? Not in a technological sense. The revolution there has already happened, in the form of the now-ubiquitous large-language models. (And note that power users could already instruct pre-existing AI models to write code that could then engage in some of these actions).
This is merely an iterative technological change whereby the AI is given permission to act and provided by its makers with the necessary web browser, phone and email interface to do so. But the implications are nevertheless revolutionary. AI can now do enormously more work for the average person.
These new consumer agents can read and monitor your email, adding meetings to your calendar and alerting you to outstanding action items. They can even send emails. They can book an appointment with your doctor or make a restaurant reservation (via email, web or even phone call). They can book an entire trip for you, shop on your behalf, haggle over your phone bill and do any number of other delegated tasks.
All of this obviously requires a high level of trust from the user, though the agents can be told not to do anything without your explicit approval. However, that micro-managing also reduces their theoretical usefulness.
In principle, consumer agents could be adopted extremely quickly. The allure of having mundane tasks handled for you is compelling.
Consumer agents also require a high degree of knowledge about their user, or else how will it know where and when to book that trip for you, or whether you prefer the pleated or unpleated slacks? Of course, after the agents have perused through your calendar and email – not just the inbox but your sent and archived emails – it is possible to imagine them getting a pretty good sense for your habits and interests.
In principle, consumer agents could be adopted extremely quickly. The allure of having mundane tasks handled for you is compelling. Software can be replicated instantaneously and be distributed rapidly (do note that there must be underlying data centre capacity, but these models should not be especially taxing).
These models are (at least initially) also free, eliminating another potential friction. The question of trust is not yet fully resolved, but one could imagine a user handing over a password or two, enjoying initial success, and then being persuaded to share their entire suite of passwords plus their credit card number.
There are also strong business uses – office workers automating a range of mundane tasks they would otherwise have to do themselves.
Who stands to win?
The households that will use these consumer agents constitute the main potential winners. If all goes well, they will save money, procure superior products, save time and organize their lives better.
If the products do help people save money by finding better deals and better products, it could be mildly deflationary, which in turn could exert a small downward influence on nominal borrowing costs as well.
The tech companies developing these consumer agents also stand to win, obviously. The ultimate monetization strategy is not yet clear but could either fold into existing subscription models or perhaps lean on advertising, especially as advertising-integrated juggernauts like Meta and Google play central roles. The consumer agents best positioned to win will be those with a deep knowledge of their users. That would again appear to give the advantage to the likes of Meta and Google (and perhaps Apple) over other frontier model-makers like OpenAI and Anthropic, and also over new entrants such as Instinct.
Main Street businesses with genuinely superior products and/or low prices but small market shares should also benefit, as the models should be good at sniffing out good deals and look past the bluster of marketing and brand names.
Who stands to lose?
However, many businesses may incrementally lose (see next chart). Those particularly reliant on marketing, brand and goodwill could be undermined by consumer agents’ relentless search for better value. Even those that do offer good value could find their margins whittled away by consumers’ increasingly frictionless ability to monitor for sales and superior pricing.
Figure 3. Diverging fortunes: agentic AI winners vs. losers
As of 10/05/2026. AI Agent Risk basket is represented by the MSXXARSK Index, AI Clawback basket is represented by the MSXXCLAW Index. Sources: Bloomberg, RBC GAM
The retail sector has attracted particular concern given that it is the most obvious place for consumer-oriented agents to grind down prices.
The telecom, media, insurance and banking sectors have also attracted concern given the potential for consumer agents to pressure fees lower and to unbundle services – to handle the complexity of identifying and negotiating telecom deals, to cancel that pesky streaming subscription you no longer want, to fight for your rejected insurance claim, or to manage the headache of keeping deposits with one bank, a mortgage with another and investments with a third.
Travel-related sectors might also suffer as consumer agents relentlessly monitor websites for the best possible airline fare or hotel deal. The scope for discriminatory pricing goes down.
A dash of skepticism
These implications are all likely directionally correct. However, with new technologies, there is always the risk of overreacting. There are several reasons not to overreact to these new consumer agents.
This is not the first time that the consumer has been given better information to make purchase decisions. The internet did the same thing. Price-shopping is already massively easier than it was 30 years ago by virtue of companies publishing their prices online, and even websites that aggregate such information. And yet brand names still receive a premium and corporate profit margins have not declined over the intervening years (though perhaps it contributed to some industry-level consolidation).
Businesses will attempt to counter the deleterious effect of consumer agents on their margins. Company-side AI will get better at haggling with consumer agents. Companies could even block consumer agents from trawling for prices or deals. Perhaps companies will stop offering sales if that cannibalizes demand for the rest of the year. Perhaps companies will come up with a different pricing model that relies less on price discrimination.
Households could also prove quite cautious around this new technology. It controls your email and calendar? You give it your credit card and passwords? It could be that AI power users will embrace this technology, but that the general public needs considerable convincing.
Bottom line
Overall, we expect consumer agents to have a genuine impact – substantially so at the company level, and modestly so at the economy-wide level. To the extent the biggest win should come in the form of a larger consumer surplus, one might posit that consumer agents may prove societally beneficial. In contrast, the implications for GDP are more mixed. The impact on the stock market may be slightly negative given mild but broad damage across a range of industries.
-EL
AI safety risks – genuine concerns, but existential risks are low
Attitudes toward AI have soured. The backlash against AI data centres (as we discussed recently) remains a hot button issue ahead of the U.S. midterm elections. But a second more existential concern has bubbled into the public consciousness in recent weeks, pertaining to whether the AI models themselves are safe for humanity.
This report examines why longstanding warnings about AI risks are growing louder, what the risk might be and how it can be addressed. It also explores what it means for the AI buildout and investors.
Our broad conclusion is that a significant effort to “pace the frontier” is ultimately unlikely, and so this is not a part of our base-case forecast. That said, we do continue to assign a 20% chance that AI ultimately underwhelms relative to current lofty expectations. This encompasses the possibility that the technology itself proves more limited than currently imagined, and that regulators opt to limit it.
Why the concern?
Warnings about the risks to humanity posed by AI (or superintelligent machines) are nothing new. In the 19th century, novelist Samuel Butler wrote about self-replicating machines developing consciousness. Computer scientists in the mid-20th century raised the possibility of recursive self-improvement allowing superintelligent machines to take over. Today’s industry leaders have been discussing existential threats from AI for more than a decade. Three years ago, they signed a public statement arguing that mitigating AI extinction risks should be prioritized alongside threats from pandemics and nuclear war.
Yet it was a fresh warning from a relatively unknown Anthropic (and former OpenAI) researcher that recently brought existential risks to the fore. Jacob Coxon resigned over concerns that leading AI labs “are racing straight to self-improving superintelligence and gambling with our lives.” He added that “the people building AI earnestly believe that it could kill us all by the end of the decade.”
Anthropic’s Alignment Science Lead Evan Hubinger fanned the flames by responding, “we really do earnestly believe AI could kill all humans,” assigning a greater than 10% probability to such an outcome within the next decade. Leaders at Anthropic and OpenAI subsequently released open letters detailing recent developments that have increased their concern about AI safety risks. They include:
Progress toward recursive self-improvement
AI systems are increasingly being used to build the next generation of AI – a form of recursive self-improvement (RSI). This is seen as a key waypoint on the path to artificial superintelligence. We’re still not at the point of fully automated coding or autonomous RSI (AI building better AI outside of human control), but Anthropic CEO Dario Amodei noted the level of RSI that is now taking place across the industry has visibly accelerated the advance of AI systems in the past few months.
A widening gap in our understanding of AI
As AI advances more rapidly, it is becoming increasingly difficult for developers to fully understand how AI systems think and make decisions. OpenAI’s Chief Scientist Jakub Pachocki explains that AI systems are grown, not designed, and their inner workings need to be studied by software engineers, just as neuroscientists study the human brain. Engineers typically rely on an agent’s chain-of-thought reasoning for insight into its thought process, but AI is demonstrating growing situation awareness – it knows when it is being monitored and might change its behaviour or hide its reasoning in a testing environment.
Growing instances of misalignment
Frontier models are generally able to follow instructions (goal alignment) but doing so safely and ethically (value alignment) remains a work in progress. Frontier labs have disclosed several troubling instances of misalignment in which AI agents pursued harmful intermediate goals not assigned by developers, acting in ways that contradicted the values taught in other settings.
A prime example is the OpenAI-Hugging Face incident earlier this year. Agents in an OpenAI test environment circumvented network restrictions, manipulated internal systems to allow communication across separate evaluations and collaborated to cheat their evaluators and cover their tracks. They hacked Hugging Face (an external platform for sharing machine learning models and data sets) in the relentless pursuit of their end goal.
While damage from the incident was limited, Anthropic’s Amodei warned that a more capable system with similar misalignment issues could cause catastrophic damage. At the current pace of development, he thinks a similar swarm of agents would be capable of taking over the internet in 6-12 months.
Other frontier labs have disclosed similar incidents in which agents have broken free from testing environments, accessed the internet without authorization, hacked into external networks and took steps to cover their tracks. OpenAI recently revealed the first known case of a rogue AI agent hacking a government website, in Australia.
What is the risk?
Anthropic’s Hubinger is not alone in his grim P(doom) assessment – that is, his subjective probability of human extinction brought about by AI. Here are some of the probabilities assigned by other experts in the field:
Figure 4. AI risk probabilities vary widely
As at 09/29/2026. Sources: RBC Assist Pro, RBC GAM
A 2023 survey of nearly 3,000 machine learning researchers found a median P(doom) of 5%, while a more recent but smaller survey found a median of around 20%. Reflecting the massive uncertainty surrounding the topic, both surveys showed estimates ranging from nearly 0-100%.
It’s worth noting that these estimates are speculative and unscientific:
They aren’t based on historical data or repeated observations.
They can’t be verified, measured or tested ahead of time.
They are entirely subjective rather than based on repeated trials or established theory (inductive or deductive reasoning).
Researchers can be a bit more scientific about the pace of AI development – when the technology might achieve a level of superintelligence that significantly increases the risk of catastrophic outcomes. Daniel Kokotajlo's AI Futures Model suggests fully automated coding of AI projects could be achievable in a year’s time, on the way to artificial superintelligence by the end of 2028. But while the model is grounded in observable trends and detailed assumptions, Kokotajlo acknowledges there is still plenty of intuition and guesswork involved.
Doomsday predictors don’t tend to specify exactly how superintelligent AI might wipe out humanity. Potential theories include AI taking over and then destroying critical infrastructure, designing and releasing bioweapons, triggering nuclear war or causing society to collapse (perhaps due to mass unemployment, misinformation, loss of trust and shared values, or erosion of human knowledge and agency).
In our view, a major but non-existential catastrophe is far more likely than AI wiping out humanity.
Alternatively, humans might simply be collateral damage in an AI system’s relentless pursuit of more benign objectives – like the “paperclip maximizer” thought experiment in which an advanced AI system consumes all resources and wipes out humanity in an effort to maximize paperclip production. These scenarios might be driven entirely by rogue AI, enabled by manipulation of or collaboration with humans, or steered predominantly by human bad actors.
In our view, a major but non-existential catastrophe is far more likely than AI wiping out humanity. Researchers at MIT recently asked 272 AI experts about the most severe risks posed by the technology. Again, the probabilities are highly subjective, but the bar for catastrophe was lowered to “harms that include the potential for more than 1 million deaths, more than $100 billion in financial losses, or comparable civilizational-scale intangible damages.” These might be described as catastrophic risks rather than existential risks.
Within a 2025-2030 time frame, experts placed the highest odds (slightly more than 20%) on catastrophe stemming from AI possessing dangerous capabilities (e.g., deception, weapons development/acquisition, persuasion and manipulation, cyber-offense, situational awareness, self-proliferation) or the use of AI systems to develop cyber or lethal weapons (autonomous, chemical, biological, nuclear, etc.).
Other severe risks include competitive dynamics (the potential to release unsafe or error-prone models in a race for AI leadership), power centralization and social inequality, and false or misleading information. Even with cost-effective measures to reduce such harms, experts still judged the risk of catastrophe to be greater than 10%.
What is being done about it?
Improving alignment and addressing monitoring challenges is critical as recursive self-improvement threatens to widen the gap between AI systems’ capabilities and engineers’ ability to understand and control them. OpenAI’s Pachocki thinks “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.”
Given that it would be corporate suicide for any frontier model developer to unilaterally slow their rate of development relative to their peers, industry leaders are now calling for a coordinated effort to “pace the frontier” – slowing the rate of AI development to give labs more time to improve monitoring and alignment; allow international coordination, regulation and governance structures to catch up; and to facilitate public debate on how AI should be used.
Anthropic’s Amodei laid out a 3-step plan:
Embed third-party evaluators with employee-like access at frontier labs (something Anthropic and OpenAI have already agreed to do).
Coordinate across companies to establish common safety standards.
Eventually coordinate globally, including with China.
Measures to slow AI development could include limiting training compute (progress in machine learning is largely driven by increasing computational power), restricting recursive self-improvement, or applying increasingly stringent safety standards based on capability benchmarks.
Leaders of other frontier labs quickly voiced their support, but the plan has drawn plenty of criticism as well. Some warn of regulatory capture, with leading AI companies aiming to cement their dominant position, raise barriers to entry and restrain progress of rivals, including cheaper open models.
Rather than slowing down, some point to the need to continue rapidly developing AI as a defence against the dangers posed by the technology itself.
The CEO of Canadian-based Cohere called out the frontier labs for cartel-like behaviour. Others think dire warnings of risks to humanity are simply a marketing ploy to talk up the labs’ capabilities ahead of potential IPOs. Critics also suggest the AI extinction debate is a red herring, distracting from more immediate AI issues like labour disruption, cybersecurity risks, concentration of power, privacy concerns and disinformation.
Rather than slowing down, some point to the need to continue rapidly developing AI as a defence against the dangers posed by the technology itself. While more powerful AI systems significantly increase cybersecurity risks, AI can also help develop new defensive tools, identify vulnerabilities and secure infrastructure against cyber threats. It can also provide real-time protection against bad actors and rogue AI agents.
Mechanisms to deactivate AI systems in a loss of control scenario are also seen as critical. There are growing calls for pre-deployment testing of new models to assess their safety. Remember, however, OpenAI-Hugging Face and similar incidents took place in a testing environment, which speaks to the need for greater cybersecurity controls at AI labs. OpenAI has committed to third-party vetting of its models earlier in the development process.
Amodei’s call for global coordination is particularly challenging given the intense AI rivalry between the U.S. and China. Indeed, Amodei says the U.S. needs to maintain or widen its lead over China by limiting the latter’s access to cutting edge chips and cracking down on the distillation of frontier models that allows (frequently Chinese) open models to nearly keep pace. Neither Trump nor Xi has expressed support for restraining the pace of AI development. In that context, the extent to which U.S. labs can pace the frontier might be dictated by how quickly Chinese labs continue to advance.
The labs are showing some willingness to police themselves . . . But this is unlikely to delay them for long.
Given President Trump’s position, meaningful action at the federal level seems unlikely in the near term. Democrats introduced a new bill that would ban superintelligence (with harsh penalties for violators), pause development of advanced AI systems until regulations and safety standards are put in place, and establish a Department of Artificial Intelligence to oversee the industry and regulate frontier models. But even with industry support, the legislation isn’t expected to go far.
In the absence of federal action, states like California, New York and Illinois are taking matters into their own hands. They’ve passed legislation that forces AI labs to implement safety and risk-mitigation frameworks, report safety incidents and test results, and submit to third-party auditing. But a patchwork of state regulations is no substitute for federal rulemaking given the potential for geographic arbitrage by AI labs.
The labs are showing some willingness to police themselves. OpenAI is pausing training of its most advanced models while it addresses security gaps. But this is unlikely to delay them for long. Developers are now facing legal action for unintentional hacks, although legal experts suggest the risk is not material at this stage as the incidents have caused limited damage.
While there is growing industry support for pacing the frontier, private sector collaboration in the absence of legislative action (not the least waiving antitrust rules to facilitate such coordination) seems unlikely given the incentives to cheat and the risks of falling behind.
What does it mean for the AI buildout?
We don’t expect a meaningful regulatory response or a coordinated private sector effort to substantially slow the pace of AI development in the near term. It might take a major (but not existential) catastrophe to force lawmakers and/or AI labs into action. Experts suggest a non-zero but relatively moderate probability of such an outcome over the next several years.
Again, we continue to assign roughly 20% odds to AI proving to be a minor general-purpose technology – either because the technology itself fails to live up to its promise, or because development is forced to slow down substantially due to safety risks or other concerns.
If we were to end up in a more regulated, slower progressing AI scenario, we would expect the following implications:
Demand for training compute would grow more slowly or perhaps even decline if development is significantly restrained. Training currently accounts for around half of AI compute, although that share is expected to shrink as growing adoption of AI tools increases demand for inference.
Less demand for training compute would lower the cost of compute and inference prices. However, prices are thought to be relatively inelastic (lower demand results in relatively modest price declines). A narrowing performance gap between frontier models and cheaper open-source competitors could put further downward pressure on inference prices.
Lower prices and, in the near term, more compute available for inference should support AI adoption. However, slower improvement in model capabilities would ultimately work in the opposite direction by limiting the tasks that can be automated by AI. The latter force might dominate.
Hyperscalers would likely slow CapEx in the face of reduced demand for training compute. That would weigh on the future earnings expectations of those companies and their suppliers, which have been key contributors to strong equity market performance. Investors would likely suffer.
Rising adoption and some further improvement in average model capabilities can support near-term productivity growth. Even current frontier models are significantly more capable than the most commonly used models. But a more gradual improvement in frontier models would limit productivity gains and the benefits accruing to AI adopters.
If a somewhat slower pace of frontier model development allows for better monitoring and alignment and reduces the risk of catastrophic outcomes, that could ultimately be positive for AI capabilities and adoption over time. A more gradual data centre buildout might also be more sustainable in the long run, although investors could suffer some near-term indigestion if the CapEx cycle slows.
-JN
The Fed and rising yields
The U.S. Federal Reserve (the Fed) raised its policy rate by 25 bps in September, the first hike in more than three years. We think this is the start of a modest mid-cycle adjustment or policy recalibration rather than a larger tightening cycle.
Chair Warsh attributed September’s unanimous decision to several factors:
ongoing signs of economic strength (suggesting financial conditions aren’t restrictive)
insufficient progress toward the Fed’s inflation objective
geopolitical developments (the renewed increase in energy prices).
Those factors, as well as a hawkish shift in Fed communications in recent months (including Warsh’s Jackson Hole speech in August) left the market widely anticipating a September hike. That said, the Fed’s first move under Warsh being a hike rather than a cut would have been a surprise earlier this year.
Warsh once again refrained from offering forward guidance on the Fed’s future policy decisions and declined to submit a projection for the dot plot (which shows the expectations of the Federal Open Market Committee (FOMC) members for the fed funds rate). But his comments left the door open to further rate increases, and among the other 18 FOMC members that all submitted dots, 16 expect at least one further rate increase this year. A plurality expects a third hike at some point.
Figure 5. Most FOMC members expect at least one more rate hike
As at 09/16/2026. Sources: U.S. Federal Reserve, Bloomberg, RBC GAM
Our base case has been that the Fed will raise its policy rate by a total of 50 bps (i.e., one further hike after September). While some have speculated that the Fed might opt for a one-and-done rate increase to establish credibility on inflation, a single move probably wouldn’t be sufficient at this point when the market is pricing in at least 3 more hikes.
There is precedent for a one-and-done rate increase, however. Alan Greenspan’s Fed hiked by 25 bps in March 1997 but held off on raising rates further when inflation started to slow dramatically. The onset of the Asian financial crisis several months later kept the Fed on the sidelines, and its next move was ultimately a series of rate cuts in the second half of 1998. In today’s context, an immediate resolution to the conflict in Iran and rapid decline in oil prices (or some unforeseen economic or financial shock) could turn September’s move into a one-off, but that’s not our base case.
A single 25 bp move doesn’t do much to change the stance of monetary policy, let alone broader financial conditions (which also includes factors like bond yields, equity markets, credit spreads, credit availability and currency moves). The Fed has a history of mid-cycle adjustments, typically cutting rates by 75 bps – enough to shift financial conditions somewhat without a major policy pivot. An increase of that magnitude would neatly unwind the insurance cuts from late last year when inflation was lower and there were concerns about the health of the labour market.
But not-too-hot inflation data, soft jobs numbers and recent comments from Fed officials suggest limited urgency for a follow-up hike in October, with the next move more likely in December. What happens after that could depend on developments in energy markets. If oil prices are coming down in early 2027, the Fed might not need to raise rates any further, in line with our base case.
While a mid-cycle adjustment is most likely in our view, there is some risk that the Fed opts for a more substantial tightening cycle. The past 7 tightening cycles going back to the mid-1980s (when the Fed started targeting the fed funds rate directly) have seen the policy rate rise by more than 300 bps on average – even the least aggressive cycle (1999-2000) saw 150 bps of tightening. Going back further to oil shocks in the 1970s and Volcker’s battle against inflation in the early 1980s makes recent tightening cycles look downright tame.
Figure 6. Tightening cycles in the last two decades have been relatively tame
As at October 1, 2026. Sources: Bloomberg, RBC GAM
What would it take for the Fed to go beyond a mid-cycle adjustment? A significant further increase in energy prices, plus clear evidence of spillover into broader consumer prices that risks destabilizing inflation expectations could force the Fed to act more aggressively. Higher inflation and inflation expectations mean nominal rates have to rise even more to ensure real rates are sufficiently restrictive. While rate hikes can’t do much to address an energy supply shock, the Fed might need to tap the brakes on the broader economy to get inflation back to target faster.
Again, that’s not our base case, although the market has started to price in some odds of more aggressive rate increases. As of late September, the futures market indicated 30% odds that the fed funds rate would be above 5% in a year’s time (i.e., 150 bps or more of cumulative rate hikes). Market pricing has retreated somewhat, but investors remain less dismissive of a 5%+ policy rate than they were a month ago (see chart below).
Figure 7. Futures pricing shows greater odds of 150 bps (or more) of interest rate hikes
As at 10/05/2026. Sources: Chicago Mercantile Exchange (CME) Group, RBC GAM
The shift in market pricing from rate cuts earlier this year to rate hikes has generally caused the yield curve to flatten. The spread between 2Y and 10Y Treasury yields was as narrow as 17 bps in September – uncomfortably close to inversion, which has historically been a key recession indicator (although it provided a false signal in 2022-24).
But unlike some “growth scare” episodes where the yield curve bull flattens on expectations of future rate cuts, the recent flattening has been accompanied by rising yields – particularly rising real yields (see the chart below). While that reflects a number of factors (as we recently discussed), higher real yields are partially due to ongoing resilience in the U.S. economy. This might suggest a higher neutral policy rate than was previously thought (consistent with rising productivity and strong demand for capital).
Figure 8. Rising Treasury yields have been driven by higher real yields
As of 10/02/2026. Sources: Bloomberg, Macrobond, RBC GAM
Treasury yields tend to track broad trends in nominal GDP growth (see the chart below). A 10-year Treasury yield above 5% isn’t particularly alarming in the context of a 6.3% annual increase in nominal GDP. To be sure, higher yields represent a tightening in financial conditions that will weigh on growth. The Fed’s financial conditions model suggests the increase in long-term yields over the past month will subtract about 0.2% from GDP growth over the next two years, if sustained.
Figure 9. Treasury yields tend to move with nominal GDP growth
As of April 2026. Sources: U.S. Bureau of Economic Analysis (BEA), U.S. Department of Treasury, Macrobond, RBC GAM
The good news is recent drivers of U.S. GDP growth aren’t particularly rate sensitive: AI builders seem unperturbed by higher borrowing costs, and consumers’ debt servicing costs tend to adjust slowly to higher interest rates given the prevalence of 30-year fixed rate mortgages. Mortgage rates above 7% will keep housing activity subdued, but as we discussed recently, the U.S. economy has performed just fine without support from that sector. Overall, we think the U.S. economy can withstand modestly higher interest rates.
-JN
Higher yields bite but boost return potential
Global sovereign bond yields have been surging as investors grapple with stubborn inflation, elevated fiscal deficits and geopolitical risks stemming from the conflict in Iran that is pushing up energy prices. The U.S. 10-year yield climbed a sizeable 50 basis points in just the past month. Government bond yields across regions have risen to levels not seen in two to three decades (see chart below).
These moves are especially significant because the yield on government bonds represents the risk-free rate available to investors and so serves as the reference benchmark against which all lending and borrowing takes place.
As a result, the pricing of other assets is likely to feel the weight of higher yields, as all opportunities available to investors must now compete against an increasingly attractive risk-free alternative. Any stabilization in yields from here, though, could support valuations and lead to favourable returns across a broad range of asset classes.
Figure 10. 10-year government bond yields have risen to levels not seen in two to three decades
As of 10/02/2026. Sources: Bloomberg, RBC GAM
Part of the increase in yields is being driven by the fact that central banks have resumed raising rates and have signaled that further hikes are likely on the horizon. These expectations are being priced into markets, as shown by the dotted line in the chart below. As of October 1, investors are pricing in just over three additional 25-basis-point rate hikes in the U.S. over the year ahead, four hikes in Canada and three more from the European Central Bank.
Of course, the path for future policy rates is highly uncertain and will depend on economic and inflation data, Much of that data will be shaped by how the Middle East conflict evolves. But for the time being, investors are penciling in a modest further tightening of monetary conditions, which will benefits savers but raises costs for borrowers.
Figure 11. Central bank policy rates are expected to rise further
As of 10/01/2026. Forecasts, shown as dotted line, are based on futures for the U.S. and OIS (Overnight Indexed Swap) forwards for other regions. Sources: Bloomberg, RBC GAM
The interest-rate and bond-yield backdrop matters for equity investors because there is a strong link between stock and bond prices, although that relationship has shifted meaningfully in recent years. Historically, and for much of the post-pandemic era when investors were more concerned about economic growth than inflation, stocks exhibited a negative correlation with bond prices (see chart below).
In other words, stock-market rallies were accompanied by declining bond values (i.e. rising bond yields). Stock-market downturns coincided with a boost in bond prices (i.e. falling bond yields) as investors sought safety during risk-off periods.
Figure 12. The relationship between stock and bond prices is changing
As of 10/01/2026. Rolling 90-day correlation computed on daily price returns. The 7-10yr U.S. Treasury ETF ticker is IEF (iShares 7-10 Year Treasury Bond ETF). Sources: Bloomberg, RBC GAM
But following the pandemic, and especially since 2022 when central banks aggressively raised rates to combat a surge in inflation, that relationship has flipped. In recent years, with investors increasingly focused on inflation rather than growth, rising yields have mostly been associated with falling stock prices. This positive correlation between equity and bond prices makes the recent backup in yields particularly consequential for equity investors.
Our models highlight the well-established historical relationship between interest rates, inflation and equity valuations. Below are three separate charts plotting the U.S. 10-year yield, the U.S. 3-month interest rate, and U.S. Consumer Price Index (CPI) inflation, respectively, against our modelled P/E (price-to-earnings) estimate for the S&P 500. Each of the charts is downward sloping, indicating that a rise in any of these variables leads to declining price-to-earnings multiples for the S&P 500, and vice versa. This relationship is intuitive, in that higher interest rates motivate investors to demand a more attractive entry point to allocate capital toward riskier assets.
Interestingly, the relationship reverses near ultra-low interest rates, which is the environment in which markets spent most of the post-financial-crisis period. In that setting, as rates exited crisis levels and moved toward something more normal, investors gained confidence in riskier assets and were willing to pay higher multiples.
At current levels, however, the conventional relationship holds once again: higher rates and elevated inflation should weigh on valuations. (That said, should the conflict in Iran find a resolution, energy prices could decline and inflation pressures may subside, allowing central banks to tighten less aggressively. Such a scenario would likely be a tailwind for both bond and equity prices.)
Figure 13. A rise in U.S. 10-year yields, the U.S. 3-month interest rate and U.S. inflation leads to declining P/E multiples for S&P 500 . . .
As of 09/30/2026. Source: RBC GAM
Looking at valuations of the S&P 500, we can see this dynamic playing out in which higher yields are leading to lower equity-market valuations. The chart below plots the S&P 500 forward P/E alongside the U.S. 10-year bond yield, inverted on the right-hand scale. The climb in yields since late 2025 has coincided with the forward P/E on the S&P 500 declining four full multiple points to 19x from a high of 23x.
Figure 14. Since late 2025, S&P 500 forward P/E dropped as 10-year bond yields rose
As of 10/01/2026. Sources: Bloomberg, RBC GAM
This deterioration in valuation is not necessarily obvious in the S&P 500 Index itself, which is hovering near record highs. That is because earnings growth has been strong and the index is being supported by a narrow set of mega-cap technology companies.
More broadly, however, the average stock has indeed been under pressure in this rising-yield environment. The New York Stock Exchange (NYSE) cumulative advance-decline line, which tallies the number of daily advancing stocks net of declining stocks, has been steadily falling since August. The line now sits at its lowest reading since March (see chart below).
More importantly, this indicator is diverging notably from the S&P 500 Index. This suggests broad deterioration in equity-market strength, even as the bellwether index is held up by a small group of mega-cap technology companies that appear less sensitive to rising interest rates.
Figure 15. The average stock has been under pressure – NYSE cumulative advance-decline line has been steadily falling since August
As of 10/01/2026. Sources: Bloomberg, RBC GAM
Recent performance confirms this divergence. The Magnificent Seven stocks have been outperforming, while equal-weight, small-cap and value-oriented strategies have lagged (see chart below). This may suggest that higher borrowing costs are disproportionately impacting companies for which the cost of capital and inflation are more critical to profitability. By contrast, the Magnificent Seven have proven their ability to grow profits at a rapid pace largely irrespective of the macroeconomic environment.
Figure 16. Magnificent Seven stocks have continued to outperform
As of 10/01/2026. Magnificent 7 represents a cap-weighted index of the following companies: NVDA, MSFT, AAPL, AMZN, GOOGL, META, TSLA. Sources: Bloomberg, RBC GAM
The bond selloff is unlikely to persist indefinitely, though, and it is worth considering what might happen once the current rise in yields is arrested. The table below lists past instances of selloffs in the U.S. 10-year bond since 1990 and how bonds and stocks performed in the 24 months thereafter. Several interesting patterns emerge across these episodes.
'Figure 17. Current U.S. 10-year Treasury selloff has so far been smaller than average of last seven bond selloffs
As of 09/30/2026. Sell-off dates were determined by peak to trough in a total return index of U.S. 10-year government bonds. High inflation deemed as annualized inflation exceeding 3.5% through the period 24m forward return is the return in the 24 months following the bond sell-off end date.
There are seven bond selloffs captured in the table, averaging a 10.7% total-return loss, whereas the current episode has registered a 7.6% loss so far. Equity valuations typically compress, especially during periods of high inflation, and stocks tend to suffer large drawdowns.
Importantly, in the periods following bond selloffs, returns for 10-year Treasuries tend to be quite strong, averaging 9.8% annualized over the subsequent 24 months. For equities, the average return following bond selloffs is also strong, at 10.4% annualized over the same horizon.
For equities, the outlook will depend on whether inflation pressures moderate without the economy tipping into recession.
But the outcome for stocks is highly dependent on what arrests the rise in yields. If bonds rallied because inflation pressures eased and the economy continued to grow, stocks delivered returns in the high teens to above 20% annualized. In these scenarios, the positive correlation between stocks and bonds held.
If, on the other hand, bonds rallied due to growth concerns, stocks struggled. In those instances, such as the tech crash and the global financial crisis, the negative correlation between stocks and bonds materialized and bonds outperformed stocks meaningfully.
Overall, the selloff in bonds is weighing on a variety of asset prices. The challenge of persistently high inflation has imposed a positive correlation between stocks and bonds, but to the downside, meaning both asset classes have been declining together.
The silver lining, however, is that the reset in yields establishes a higher reference rate from which future returns will be measured. This effectively boosts the return potential across all asset classes.
Meanwhile, strong earnings growth has largely offset falling equity valuations. Moreover, once the bond selloff ultimately runs its course, forward returns for fixed income have historically been quite strong. For equities, the outlook will depend on whether inflation pressures moderate without the economy tipping into recession.
The range of outcomes for stocks is wide, but for bonds, the rise in yields offers investors a higher guaranteed starting point, and one that has improved meaningfully in recent months.
-ES
Distilling macro themes from earnings commentary
Another quarterly earnings season has wrapped up. Once again, we used our in-house AI tools to analyze S&P 500 earnings call transcripts for insights into what’s on the minds of management teams at the biggest American companies. AI remained a central theme. Higher energy prices were an ongoing source of cost pressure. There was more discussion of rising interest rates and defence spending, and a mixed tone on consumers.
Overall, the Q3 earnings discussion aligned with many of the themes we have been highlighting in recent reports and is consistent with our cautious optimism that the U.S. economy will remain resilient in the face of several headwinds.
AI remained a hot topic across several industries with total mentions more than doubling compared to the same quarter a year ago. There were several evolving dimensions to the AI conversation:
A range of sectors including tech, utilities, industrials, energy, materials, communication services and real estate were bullish on demand related to the AI buildout. Companies noted expanding order books; demand for electricity, gas, equipment and building materials; and increasing construction activity, leased capacity and network traffic.
That said, companies in several of those sectors also expressed caution regarding constraints to the data centre buildout including memory and foundry capacity, backlogs for turbines and electrical equipment, skilled trades shortages, and local zoning and planning permission challenges. Some sectors (including consumer discretionary and health care) also noted margin headwinds from higher prices and scarce supply of memory and electronics.
Many sectors were increasingly constructive on AI as a productivity tailwind, led by tech and financial services which have seen the most widespread AI adoption. Productivity gains were generally framed as allowing output to expand rather than shrinking headcount. Indeed, the number of job cuts in Q3 attributed to AI was the lowest in a year (see chart below).
Figure 18. Job cuts attributed to AI in Q3 were lowest in past year
As of Q3 2026. Sources: Challenger, Gray & Christmas, Inc., Macrobond, RBC GAM
Beyond AI, several other interesting topics were raised:
Several sectors were cautious about rising fuel costs and freight charges, particularly consumer discretionary and industrials. Management teams generally described cost headwinds as building, consistent with the renewed increase in oil prices and even sharper gains in refined product prices during the quarter. Companies were mixed on the extent to which higher input costs are being passed along to output prices. There was relatively limited evidence of demand destruction due to higher prices, outside of consumer discretionary where travel demand was negatively impacted.
Several sectors raised the issue of higher debt servicing costs as interest rates continued to climb. However, the tone was not particularly anxious and access to credit does not appear to be a constraint. Higher mortgage rates (now above 7% and up more than 100 bps year-to-date) also contributed to a bearish tone on housing, particularly in the materials sector. However, some real estate companies noted affordability challenges are improving the outlook for rental demand.
Tariff discussions were much less prevalent than a year ago and shifted from a cost headwind to a refund tailwind, although that impulse is already fading. As of mid-September, U.S. Customs had issued $122 billion in refunds, about ¾ of the total collected under IEEPA (International Emergency Economic Powers Act). Some companies pointed to refunds as temporarily supporting margins while others are lowering prices or passing rebates along to vendors.
Commentary on the consumer was mixed. Consumer discretionary companies noted households are under pressure from rising gasoline prices, but demand is holding up, as is shopping traffic according to real estate firms. Financial services companies said demand was robust across income groups despite higher inflation, with tax refunds and a healthy labour market seen as supportive. There was limited discussion of unemployment, delinquency or payment stress.
While AI is dominating the CapEx cycle, tech and industrials described defence spending as another tailwind. The conflict in Iran is supporting near-term demand, and as we’ve discussed in the past, many countries are ramping up their defence spending in an increasingly dangerous world where the U.S. is becoming a less reliable security partner.
-JN
Quick hits
U.S. and China summit – only minor action
The leaders of the world’s two economic and military superpowers – the U.S. and China – met in Washington, D.C. from September 23 to 25. There were no major breakthroughs, though several minor items were addressed:
The two countries agreed to delay the expiry of their current tariff détente by another two months, from November 10, 2026 to January 10, 2027. Absent this, the tariff rate each applies to the other would have increased by a whopping 24 percentage points (from 10% to 34% on many products). Of course, that cliff now looms in early 2027 – there was no long-term solution.
As a show of goodwill, each country promised to lower tariffs on US$30 billion of imported goods from the other. That represents approximately 10% of the trade conducted between the two parties. China is expected to reduce tariffs on U.S. corn, wheat, meat, dairy products, seafood, timber, coal, cosmetics and medical devices (though not soybeans). The U.S. is expected to lower tariffs on Chinese small appliances, household goods, toys and games, and sporting equipment. How far the tariffs will fall on these products was not defined, nor was the implementation date.
China promised to buy 10 million tonnes of U.S. coal per year in 2027 and 2028. China imported more than this in 2024, so it was once the norm. But coal imports collapsed in 2025 and 2026 due to China’s imposition of retaliatory tariffs. Given that China pledged to lower tariffs on U.S. coal as per above, this revival may happen organically. Conversely, note that China failed to live up to a range of purchase commitments it made in the late 2010s.
The two countries will meet again in November and then again in December at international conferences.
5000-dollar cheques – highly unlikely
U.S. midterm elections are now less than a month away (November 3). As long expected, the House of Representatives appears set to flip to the Democrats, with a 92% probability as per Polymarket. More recently, Senate expectations have also shifted, with a 66% chance it now also turns toward the Democrats (see next chart).
Figure 19. Current U.S. 10-year Treasury selloff has so far been smaller than the average of last seven bond selloffs
As of 10/05/2026. Sources: Polymarket, Bloomberg, RBC GAM
For those two reasons, President Trump’s September 9 pledge to pay a “Trump dividend” of US$5,000 per adult if the Republicans manage to maintain their two Congressional majorities is unlikely to be implemented.
But for the sake of completeness, let us nevertheless spend a moment on it.
The offer is obviously quite alluring to the average voter. It constitutes a large sum of money, and doubly so given the prospect of two such payments for most families. It also comes at a time of challenging pocketbook economics, with a 7.36% 30-year mortgage rate and gas at $4.36 per gallon.
But the reality is that, even in an unlikely Republican sweep scenario, it would be exceedingly difficult to deliver on the $5,000 promise.
Collectively, the initiative would cost the government an incredible $1.3 trillion. Further, it would arrive at a time when the fiscal deficit is already exceptionally large (see next chart), with the bond market already groaning (and demanding higher yields).
Figure 20. U.S. fiscal deficit is chronically large
As of August 2026. Shaded area represents recession. Sources: U.S. Department of Treasury, Macrobond, RBC GAM
If pandemic payments provide any guide, about a third of any such stimulus would be spent on consumption. The resulting $400 billion-plus that would come gushing into the economy would constitute potentially problematic fiscal stimulus at a time when inflation is already too high and the economy is already operating around its potential. The Fed would almost certainly have to raise rates further to offset the effect of the stimulus. This would defeat the purpose.
Critically, a Republican Congress would be unlikely to muster enough support to deliver on the promise, as there has been substantial Republican opposition to it.
In conclusion, the $5,000 cheques are an interesting thought exercise and may serve as an inducement for some voters. But for several reasons, it is highly unlikely the stimulus will be implemented.
-EL
Greenland deal – little changes
On September 22, a U.S.-Denmark-Greenland security deal was formally signed at the United Nations in New York City. This comes after almost two years of pressure from the U.S., including the White House’s assertion that “the ownership and control of Greenland is an absolute necessity.”
This deal reduces one source of U.S.-related geopolitical tensions. But how much actually changes relative to the pre-existing treaty that was signed in 1951?
The U.S. now has permission to expand its existing military base and to construct two more in Greenland. This is material, though it should be noted that the U.S. had as many as 17 bases/military installations in the past (some quite small). One of the new bases will be brand new, while the other will be constructed on the site of a former base. New bases were also permitted under the prior treaty and also subject to Greenland’s approval.
So, the relevant development on this subject is Greenland’s approval of this specific expansion plan, not the new treaty versus the old one.
To summarize, the deal does significantly obstruct the likes of China and Russia from accessing Greenland, in both a military and an economic capacity.
The new deal bars any non-NATO military presence in Greenland, so China and Russia would be blocked. Of course, Denmark and Greenland were never likely to approve such overtures. But the door is still open for other NATO members – subject to Greenland’s approval, of course.
The U.S. can now block foreign investments into Greenland (such as resource deals) that originate from outside NATO or the EU. Of course, that means most of the developed world can invest freely within Greenland. And whether the suitor is the U.S or otherwise, Greenland could still block investments.
To summarize, the deal does significantly obstruct the likes of China and Russia from accessing Greenland, in both a military and an economic capacity. The U.S. has a privileged position by virtue of its three approved bases, but not an exclusive military right relative to other NATO members. The bases may help to monitor Russia’s Arctic capabilities, limit its geographic claims, provide early missile warning and improve access to polar-orbiting satellites.
There seems to be no special resource access for the U.S. relative to other NATO or EU members.
This outcome further informs the revived Monroe Doctrine, through which the U.S. claims a privileged position in the Americas. In practice, this has so far ranged from the hostile U.S. takeover of Venezuela to this mutually agreed-upon deal – albeit under some pressure – with Greenland. The fate of Cuba is still pending, and the U.S. has in the past voiced disapproval of or ambitions for Colombia, Panama and Canada. It remains a dangerous world.
-EL
Canada’s corporate tax cut – big deal
The federal government recently announced a significant change in corporate taxes to incentivize new business investment. The Productivity Mega Deduction will permanently allow immediate expensing of a broader range of capital assets, with 65% of business CapEx now covered compared with 15% previously. Immediate expensing improves cash flow and allows faster recovery of investment costs. This ultimately lowers a company’s tax burden via the time value of money.
According to the government’s calculations, the expanded deduction lowers the marginal effective tax rate (METR) on new investment to just 6.4%, well below the 16.9% rate in the U.S. and the 26% average in other G7 countries (see chart below). The U.S. rate was lowered by bonus depreciation measures in last year’s One Big Beautiful Bill Act (OBBBA).
As a reminder, there are different ways of calculating METR, and other measures have historically suggested Canada has less of an advantage than the government claims. Other tax rates, including the average effective tax rate and statutory corporate tax rate, are also relevant. See here for our previous discussion of this issue.
Still, regardless of the measure, Canadian business taxation has just taken a huge leap forward and renders the country quite competitive versus its peers.
Figure 21. Canada has lowest marginal effective tax rate in the G7
As at 09/15/2026. The marginal effective tax rate (METR) is the tax imposed on an additional dollar of business investment U.S. tax cut reflects OBBBA . Canada tax cut reflects Budget 2025, Spring Economic Update 2026 and recent Productivity Mega Deduction announcements. Sources: Government of Canada, RBC GAM
The roughly two-thirds of CapEx now covered by immediate expensing in Canada is broader than the 45-55% coverage offered by the U.S.’s OBBBA. Canada’s deduction now includes infrastructure assets like pipelines (other than for natural gas distribution), railways, bridges, roads and fibre optic cable, as well as mining property and development expenses – most of which is not eligible under the U.S. OBBBA. Most buildings aren’t covered, although manufacturing and processing facilities are temporarily covered by the Super-Deduction that will be phased out starting in 2030.
The government estimates the tax cut will cost $36 billion over 5 years but will add up to $22 billion to average annual economic output over 10 years (0.5-0.6% of GDP). Measures that increase business investment (and ultimately productivity) typically have a high fiscal multiplier, although such an add to growth will depend on businesses taking advantage of CapEx incentives amid ongoing trade policy uncertainty.
As we discussed recently, competitive tax policy is just one factor that drives business investment decisions. But Canada is making progress in other areas as well – particularly a more business-friendly regulatory environment. Last year, foreign direct investment into the country rose to its highest level (as a share of GDP) in a decade. At the recent Canada Investment Summit, institutional investors announced sizeable investment pledges and Canadian banks made significant lending commitments.
In the past week alone, a $30 billion expansion of the Liquefied Natural Gas (LNG) export terminal in Kitimat, BC – which will eventually make it one of the largest such facilities in the world – was given the go-ahead. An oil pipeline to Delta, BC was designated a project of national interest, allowing it to go through an accelerated approval process.
Expanding Canada’s export infrastructure will allow the country to better leverage its natural resource endowment as trading partners increasingly look for safe, dependable suppliers of key resources. Canada’s policy mix is becoming quite supportive for the country’s long-term growth prospects.
-JN
Canada’s population is actually growing
It turns out Canada’s population isn’t shrinking after all. Previously reported data suggested net outflows of non-permanent residents (NPRs) were causing Canada’s population to decline outright for the first time in the country’s history. But StatCan warned of potential upward revisions given a growing number of NPR permit extensions and longer processing times. Indeed, when the annual figures were released in September, the estimate of NPRs was revised up by 240,000, and the overall population increased by 0.5% year-over-year as of July 1.
This has several implications. We thought Canada’s relatively sluggish GDP growth was partially explained by a declining population, with per capita measures looking less bad. But with the upward revision to historical population growth (which shouldn’t impact GDP estimates), per capita growth looks slower, and productivity growth was likely worse over the past year or two.
The revisions also mean the federal government is further from reaching its target of shrinking the NPR share of the population to slightly less than 5% by the end of 2027. The NPR share is now estimated at 6.7%, 0.5% higher than previously thought and only 0.5% lower than the October 2024 peak. We are skeptical that the government will hit its target on time, which should result in more even but slow-for-longer population growth in the next few years (see chart below).
The upshot is that potential growth in Canada’s economy will remain relatively sluggish (1.0-1.5%) absent a notable pickup in productivity growth (which we do forecast).
Figure 22. Canada’s population growth is expected to remain subdued
As of 2026. Sources: Statistics Canada, Macrobond, Parliamentary Budget Officer, RBC GAM
Canada joins EU? Mostly symbolic
The Canada-European Union (EU) love-fest continues. European Commission President von der Leyen has proposed “associate” EU membership for Canada given the common values of the two regions and the imperative to foster tighter bonds between like-minded middle powers as the international order becomes fractured.
What does this mean, exactly?
Canada already has a free-trade agreement with the EU called CETA that took effect in 2017. Thus, the most obvious opportunity has already been implemented. There is also the relatively free flow of capital between the jurisdictions, but only limited free exchange of services and no real freedom of movement beyond tourism. As such, there is certainly theoretical scope for a deeper relationship.
But expectations need to be kept in check. The EU cannot just give Canada freedom of movement and other privileges. It already has a menu of different levels of economic integration, and for every privilege there is an obligation (see next graphic). Countries in the European Economic Area such as Norway enjoy nearly full economic integration with the EU, but in exchange must align with relevant EU regulations and also contribute to the EU budget. Switzerland does it a bit differently, but with a similar end result.
There are then countries that are partially integrated, such as Türkiye and Ukraine. The specifics vary by country, but they enjoy fairly deep trade integration and get to participate in some EU programs in exchange for adopting some EU rules.
Figure 23. EU integration: for every privilege there is an obligation
As at 10/05/2026. Source: RBC GAM
Recall that Canada just objected to U.S. demands to limit Canada’s trade dealings with other nations. In contrast, Türkiye, which has a customs union with the EU, accepted precisely this restriction – it aligns with the EU on all trade deals. Canada would presumably not be willing to do this for the EU if it will not for the U.S. Given fresh Canadian sensitivities about ceding sovereignty, it would also be difficult to adopt many EU rules – a prerequisite for significantly deeper economic integration.
Joining the EU as a full member was never on the table, but entering the European Economic Area is also therefore unlikely. Even partial integration of the sort embraced by Türkiye and Ukraine might be a bridge too far.
It must also be noted that the President of the European Commission doesn’t just get to decide who integrates, or how. European Parliament would have to approve, as would all 27 EU member nations.
For context as to how difficult that would be, Canada’s CETA trade deal, while “provisionally” in effect for nearly a decade, has still not been ratified by 10 of the 27 nations! It exists in a political limbo whereby so long as the countries that have rejected it don’t formally inform the EU of their rejection, it remains in effect. This is precarious and conveys the political complexity of trying to secure an even deeper connection.
Instead, we assume that any associate EU membership is primarily symbolic, but reflects a handful of targeted efforts to cooperate that exist outside of the standard Eurozone vs. EU vs. European Economic Area rubric:
As of March 2026, the EU and Canada have formally commenced negotiations for a digital trade agreement that would complement CETA.
The EU and Canada have agreed to partner on digital cooperation – with relevance to AI issues and technology more generally.
The two regions have already agreed to a new defence procurement agreement, with scope for Canadian firms to sell to European militaries and for both parties to collaborate on joint purchases.
Canada was selected to be headquarters for the Defence, Security and Resilience Bank (DSRB). Membership is presently just Canada plus several small European countries, but it has aspirations of becoming a World Bank-type entity to help finance military production and security projects.
The two parties have formed a critical minerals strategy – a production and supply-chain alliance to ensure sufficient supplies of such materials.
Canada and the EU already have a research-funding collaboration, with plans to deepen this in 2028.
Prime Minister Carney has proposed that Canada join Erasmus+, a program that encourages cross-border post-secondary studies across a range of countries.
There may also be discussions about permitting greater youth mobility between Canada and the EU – allowing them not just to study but also to work on either side of the Atlantic.
Other areas of mutual interest include energy supplies and Arctic security.
To conclude, Canada is unlikely to join the EU or European Economic Area. Even partial integration is likely too large a step. But there are several opportunities to deepen ties in limited ways, listed above, and there is a good chance of many of these being implemented to the mutual benefit of both parties.
-EL