The Great Hedge: Drugs, Debt, & the Direction AI is Going Everybody’s Waiting on

Regulators, central bankers, and a Shanghai drug-discovery lab spent the week writing the terms of a bet they can’t grade yet.

In Zurich, Washington, and Seoul, the people who set the price of money are making the same admission a Chinese biotech CEO just made about the price of a new drug: nobody knows yet whether artificial intelligence is the thing that saves the timeline or the thing that blows up the budget.

A Bet Nobody Can Read

Insurance doesn’t wait for an Act of God to name itself. It prices the sky before the storm arrives, on the assumption that not-knowing is itself a cost, one somebody has to carry. That is what regulators, bond markets, and a handful of finance ministries did this week without quite saying so out loud.

The Federal Reserve’s Kevin Warsh has nodded to an AI-fueled productivity boom as one route back to target inflation, while Swiss National Bank board member Petra Tschudin said the technology could push prices up in the short term and down only much later, if at all.

Neither is wrong, exactly. Neither is right, either.

What connects a Shanghai drug lab, an Indian outsourcing contract, and a Seoul finance ministry memo (in an abstract sense) is not a shared answer, it’s a shared refusal to wait for one. Every institution in this chain of stories is pricing a variable whose sign it cannot confirm, which means the real story is not what AI is doing to the economy. It’s what the economy is doing while it waits to find out.

The Speed Trade

Insilico Medicine, the Hong Kong–listed drug developer, says it has cut the time to a developmental candidate to about one year, and in its best case to nine months, by pairing frontier AI research in Montreal and Abu Dhabi with clinical execution in Shanghai.

“It usually takes about 4.5 years to get to a drug developmental candidate using the traditional approach,” CEO Alex Zhavoronkov told Reuters at the company’s Shanghai R&D facility. “But if a pharma company has a research lab in China… they cut two years of time. When you combine frontier AI that is proven to work experimentally with the power of China, you can significantly accelerate that.” 90% of Insilico’s revenue still comes from Western pharma partners, and its lead AI-designed drug has only reached Phase II trials, so the speed claim is a pipeline efficiency, not yet a market one. But a senior Pfizer executive made a nearly identical claim about China’s broader ecosystem last month, saying clinical development there runs three times faster and at roughly half the cost of Europe, which is part of why the US moved in June to speed up its own drug-development guidance.

“In my organization with 400 people, I can probably displace 40% easily on the software side.”

— Alex Zhavoronkov, CEO, Insilico Medicine

Regulators are watching the same clock from the writing-the-rulebook side

The FDA and the European Medicines Agency (EMA) jointly published ten guiding principles for AI in drug development in January, lean on principle 2 (“risk-based approach with proportionate validation”) and principle 9 (“scheduled monitoring and periodic re-evaluation”) rather than fixed thresholds, precisely because the technology moves faster than a rulebook can be revised.

When industry stakeholders pushed EMA for more concrete detail at a February 4 meeting, including templates, checklists, and clearer boundaries between AI decision-support tools and regulated medical-device software, regulators pushed back on the premise itself: “It is important to manage expectations about how detailed regulatory guidance on AI can be… many issues depend on context, therefore very detailed rules may not be feasible.”

That’s not a deferral, it’s a design choice consistent with the ten-point framework: EMA’s concurrent workstreams — a concept paper on AI in clinical development, a Q&A-style pharmacovigilance guidance built jointly with its safety committee, and a manufacturing annex revised after roughly 1,300 public comments — are all context-specific instruments rather than a single detailed code, which is the risk-based principle applied to the guidance-writing process itself. No one has priced which way it breaks.

Repricing the Human Ledger

One of the clearest places this is already showing up in contracts, not just guidance, is India’s $315 billion IT services industry.

TCS, India’s largest outsourcing firm, now ties about 80% of its finance and HR services contracts to performance outcomes rather than hours billed, roughly double the share from before AI went mainstream in late 2023.

Persistent Systems’ CEO says clients are demanding the same work for 25% to 30% less while expecting faster delivery. The same week, a Naukri jobs-portal survey showed AI-specific hiring in India’s IT sector rising 16% year-on-year even as overall IT hiring fell 3%; very suggestive evidence that firms are reallocating headcount toward AI roles rather than simply cutting it, though TCS itself cut more than 12,000 jobs last year.

“The pyramid model is gone,” former Infosys CFO V. Balakrishnan said. “With coding agents, we no longer need basic coding.”

The Money’s Already Moving

Capital markets are pricing the same uncertainty at a much larger scale.

US 30-year real yields have climbed to near 18-year highs, driven in part by roughly $220 billion in bonds issued this year by AI hyperscalers including Alphabet, Amazon, and Meta, more than double their combined 2025 total. “There’s a competition for capital which is relatively unprecedented in recent times,” BlackRock’s Vivek Paul said.

South Korea is trying to get ahead of a related version of the same problem. Its budget ministry plans a “Future Response Fund” that could exceed 100 trillion won ($72 billion), built from tax revenue above trend, largely windfall from Samsung and SK hynix chip earnings, to be spent on AI investment and youth programs while insulating the budget from the boom’s eventual reversal.

Capital markets are pricing the same uncertainty at a much larger scale.

US 30-year real yields have climbed to near 18-year highs, driven in part by roughly $220 billion in bonds issued this year by AI hyperscalers including Alphabet, Amazon, and Meta, more than double their combined 2025 total. “There’s a competition for capital which is relatively unprecedented in recent times,” BlackRock’s Vivek Paul said.

“Because of things like the AI build-out ramping ever up, that capital scarcity dynamic is accelerating and you’re seeing that play out in bond yields.”

— Vivek Paul, Global Head of Portfolio Research, BlackRock

Not everyone reads it as a warning sign yet, stocks are still at record highs, but Satori Insights founder Matt King wrote that he expects real yields to “continue rising until they choke off the borrowing which has been driving them, and the rotation into risk which has been fueling the equity rally.”

The Fed is having a more pointed version of the same argument about the sign of the effect, not just its size. In a Reuters Open Interest column published August 17, columnist Jamie McGeever laid out the three routes Warsh has nodded to for returning inflation to target: raising interest rates, shrinking the Fed’s balance sheet, or “relying on an AI-fueled productivity boom to cool price pressures.”

Rate hikes, McGeever argued, are the least palatable of the three but by far the most likely to work, while the other two are “poor substitutes.” The productivity route in particular has a timing problem layered on top of an evidence problem. Monetary policy itself already works with a lag once thought to run 12 to 24 months, which Fed Governor Christopher Waller now estimates at closer to 9 to 12. The transmission of an AI productivity gain into actual price relief, McGeever writes, “take[s] even longer. That’s if they exist at all.”

On the evidence side, the San Francisco Fed’s utilization-adjusted Total Factor Productivity index has turned negative on a rolling four-quarter basis, with its economists noting in May that “broader efficiency gains (from AI) remain unrealized so far.” John Silvia, CEO of Dynamic Economic Strategy, put the underlying problem most bluntly: “When it comes to the real world, and I’m dealing with the actual data from 1982 to now, I cannot get a statistically significant relationship between” productivity growth and inflation.

Swiss National Bank’s Tschudin, in an interview published the same week in Finanz und Wirtschaft, described a mechanism for why AI might do the opposite of what Warsh is hoping for, at least at first. “Investment flows are being partly redirected, which can mean adjustments and difficulties for the rest of the economy,” she said. “Shortages can occur, for example with chips, causing prices to rise. In the short or medium term, therefore, upward inflationary pressure can also arise.”

She didn’t rule out the disinflationary case Warsh is banking on. Longer term, AI-driven productivity could make goods cheaper, but she was skeptical it resolves cleanly: because inflation is measured annually, a genuine deflationary effect would need to repeat “regularly,” and “productivity gains as such are not a new phenomenon. They do not, by themselves, lead an economy into structural deflation.”

The IMF’s new chief economist, Silvana Tenreyro, made the same point independently in research published by Bank of England staff two days earlier: even if AI does boost productivity, that alone may not lower inflation.

South Korea’s central bank is living the tension in real time rather than debating it in the abstract. Its July 16 rate-hike statement cited AI investment in the same document as a reason for both confidence and caution: “the global economy is expected to continue its moderate growth trend, driven by robust AI investment,” reads one line, while a later line in the same release notes that “stock prices underwent a sizable correction amid heightened volatility due to growing concerns over AI investment.”

Seven of seven board members still voted to raise the base rate to 2.75%. Seoul’s Ministry of Planning and Budget is meanwhile trying to insulate its own budget from the reverse of that same volatility: its proposed Future Response Fund, potentially exceeding 100 trillion won ($72 billion), would bank the tax windfall from Samsung and SK Hynix’s chip earnings against the day the AI investment cycle turns. No one has priced which way it breaks.

---

---

Elsewhere in Asia: a Related Bet That Already Picked a Direction

Everything in this piece so far describes institutions hedging against a variable whose sign they can’t yet call, whose payoffs are still hazy on the horizon. It’s worth flagging, then, a different but adjacent story we’ve been watching in Asia’s capital markets, because it may be gesturing at some of the same underlying mechanisms, even though it isn’t really the same kind of case.

One line of reporting we’ve followed appears to suggest that compute itself is increasingly being treated as a financeable asset class in parts of Asia’s markets, something closer to a financial instrument than a hardware input. There’s Nvidia’s roughly $500 billion financing platform and what’s being called “computing banks.”

“These new models essentially financialize computing power,” is how one economist, Pan Helin, put it. There’s also a fairly stark divergence in how that’s playing out: SK hynix’s Seoul listing reportedly drew seven times oversubscription at $26.5 billion, while a comparable Taiwanese peer, Unimicron, priced at a discount around $1.4 billion in the same window. If that gap holds up, it would appear the market isn’t pricing this evenly across the sector — precisely how Reflexivity President Giuseppe Sette characterized it.

In a loose parallel, this Asian capital-markets story doesn’t read like an open question the way the drug-regulation, IT-contract, and monetary-policy cases do. The direction here already looks fairly set: capital is flowing toward treating compute as a tradeable asset, not hedging over whether it should. Whether that’s actually the same underlying mechanism at an earlier or more advanced stage, or just a different market doing something that happens to look structurally similar, remains to be seen.

---

Generative AI Transparency:

This article was written primarily with generative AI, specifically SupraGraphos’ A.C.E. News Module. Reviewed with human post-editing, all sources and claims are confirmed as of the time of writing.