Opinions

Residual Value Guarantees (RGVs) have emerged as a new tool in AI debt financing in the past few months. But RVGs did not originate in AI finance; they have long been used in aircraft leasing, equipment leasing, and auto leasing. Their re-emergence in AI debt reflects a common problem: lenders are being asked to finance assets whose future value is unusually hard to forecast, and a guarantee from a well-capitalized sponsor is one of the few tools available to bridge that uncertainty. Rental rates for Nvidia's H100 chip are reported to have fallen from roughly $8 per hour in early 2024 to $2–3 per hour by late 2025, a decline of 60–70 percent in about eighteen months. High-performance GPUs are also generally understood to face a two-to-three-year obsolescence cycle, while the debt used to finance them commonly carries three-to-five-year (and in some data center deals, twenty-year) terms. That mismatch between the useful economic life of the asset and the tenor of the financing is the central problem that residual value guarantees, along with related credit backstops, are designed to address. In its simplest form, an RVG obligates the guarantor to pay the difference between an asset's actual fair value at a defined point — typically lease expiration, non-renewal, or early termination — and a pre-agreed floor value, if the actual value falls short. The guarantee does not eliminate the underlying depreciation risk; it reallocates who bears it, shifting exposure from the debt or equity investors in a financing vehicle back toward the corporate sponsor that benefits from using the asset. For sponsors, an RVG is attractive because it unlocks financing that would otherwise be difficult to raise. For lenders and equity investors in the SPV, the guarantee substitutes the sponsor's corporate credit — typically investment-grade — for an independent judgment about what a data center or a fleet of GPUs will be worth years from now, which is precisely the judgment that is hardest to make reliably in this asset class. We see following risks in RGVs: 1. Depreciation uncertainty 2. concentration risk 3. disclosure & balance sheet treatment 4. pricing with out a benchmark. Current accounting standards treat residual value guarantee as a contingent liability: because payment depends on a future asset value falling short of a threshold, which may or may not happen. Companies commonly disclose the maximum exposure in the notes to their financial statements without recording it as a liability on the balance sheet unless and until a payout becomes probable and estimable. To summarise we believe RGVs are procyclical in nature & tail risks are elevated given off balance sheet nature of RGVs. No wonder the Nvidia 5-year CDS has moved up sharply to 75 points currently from 45 points in end June. Same applies to Broadcom 5-year CDS which has moved up sharply to 99 points currently from 50 points in end June.

The sell-off in AI equities has received more press, but AI-related debt indices have substantially lagged too. The JPM high Grade High Performance Compute (HPC) index has widened 38bp since the start of June to 208bp. For reference, the JPM HY BB Index trades at 201bp. JPM High Yield HPC index has widened 153bp in the last month to 418bp. Despite the growing pains, the universe keeps getting bigger. The combined hyperscaler and data center universe across the high grade and high yield indices encompasses 31 different issuers with over $576 billion in outstanding bonds as well as over $5 billion and growing of leveraged loans. Technology has just surpassed US Banks as the largest sector in JULI at 11.6% of the index and has accounted for 33% of net HG issuance YTD so the market wide impact of hyperscaler spread moves will continue to grow. At ~$870bn and +77% y/y, AI capex is no longer a sector story — it is the US business cycle, and also it’s largest fragility. Hyperscaler free cash flow has hit zero — a choice at the top of the stack, a necessity beneath it. Operating cash flow ~$770bn has met capex ~$750bn — the buffer is gone; Alphabet's FCF is negative for the first time since its 2004 IPO. But the first domino is not the top of the stack — it is the levered tier beneath it. The levered tier is Oracle which spends 83% of revenue on capex with $167bn of debt one notch above junk; neoclouds spend 2–8x revenue. The long end is the pressure valve: 30Y real yields at 25-year highs as AI paper collides with a $2.1trn federal interest bill. When the large IG bond supply hits along with the large -dated UST supply, spreads will have to compensate for investor demand. Hence the IG as well as HY spreads especially in the HPC sector might expand significantly. Higher spreads mean higher interest cost, implies lower free cash flows implies lower valuations implies P/E derating implies lower equity valuations implies pressure to reduce capex guidance implies higher spreads and the cycle continues. With UST yields staying elevated, hyperscalers bond absolute levels as well as spreads might be under pressure for some time now. Unless the macro data turns, this implies the end of AI capex cycle is not very far. And this time it won’t be just equities which will suffer but private credits/BDCs/funding banks might suffer more.

Last week, the Trump administration announced new tariffs under Section 301 to replace expiring Section 122 universal 10% tariffs. A transition to Section 301 from Section 122 tariffs will likely increase the average effective tariff rate by about 1pp to 8%. We continue to expect that increased focus on affordability issues and elevated inflation will likely prevent US tariff policy from escalating drastically going forward. Our expectation for terminal average effective tariff rate is at 8-9% range. Also, details of the new tariff announcements indicate a rising share of exempted items. We do not view last week's announcements as a sign of escalation in US tariff policy. . Most trade law experts view the section 301 tariffs to be on somewhat firmer legal ground, but that likely will not preclude legal challenges to the forced-labor and excess capacity justifications offered for the most recent actions. To summarise, we do not expect any material change in US tariffs now since mid-term elections are very close. Also, the rise in gasoline prices above $4 levels implies a constrained US consumer spending capacity, thus making tariff related price increases a hot political issue which Trump might want to avoid at least till Nov’26. The above view adds to our conviction that Fed is likely to remain on hold for REMCY26. We are currently received half risk at 4.15 in 1yr-1yr US SOFR and another half risk waiting to get received at 4.25. This week’s high was 4.23 before it cooled down to close at 4.20. We have a profit target of 3.9 and stop loss at 4.4 for this trade.

The entire US AI capex juggernaut has been threatened by the release of Chinese AI firm Moonshot's Kimi K3 release last week. The Kimi K3 AI model delivers performance on industry benchmarks that rivals top-tier offerings from OpenAI and Anthropic PBC at cheaper costs. Even from a technical perspective, Kimi K3 contains 2.8 trillion total parameters, making it one of the largest open-weight AI models ever released. Though, Claude Fable 5 & Mythos 5 has 6 trillion parameters along with GPT 5.1 Sol between 2-3 trillion parameters, it Kimi K3's ability to do long horizon work rather than simply answering questions which makes it second to Claude Fable 5 at one third the cost. Claude Fable 5 costs $10 per million input tokens and $50 per million output tokens. GPT- 5.6 sol costs $5 per million input tokens & $30 per million output tokens. Kimi K3’ costs $3 per million input tokens & $15 per million output tokens. Kimi K3's native image and video understanding also allows it to revise interfaces, games and other digital content after inspecting its own output. At launch, Moonshot reports 93.5% on GPQA Diamond, the best open-weight score ever published on that benchmark, 88.3% on Terminal-Bench 2.1, and a record 91.2% on Browse Comp for web agents. On Artificial Analysis' private long-horizon knowledge-work evaluation, K3 reaches an Elo of 1547, a massive jump from K2.6 and behind only Claude Fable 5. K3 slots in as the best open-weight agent, just behind Claude Fable 5 on reasoning depth and roughly even with GPT-5.6 Sol on terminal work, at a price between them. Hence Moonshot’s Kimi K3 threatens the very base of US AI Capex. Moonshot looks like a viable alternative, that could upend the business case for its US AI competitors. It might threaten the revenue model and eventually put a doubt in the mind of investors if current AI capex run rate in US is sustainable. Chinese models have already priced their services aggressively, undercutting US rivals. The weighted average cost of performing a standardized intelligence task using Kimi 2.6 or DeepSeek’s V4 Flash model ranges from 33 cents to just 2 cents, compared with $2.75 for the same task on Anthropic’s Claude Fable 5, according to benchmarking site Artificial Analysis. Moonshot’s latest offering is priced at a premium by comparison. Kimi K3’s cost on the same ranking is estimated at 95 cents. Kimi does not have to be the world's single best model to upend the US AI dream run. For companies, governments and developers, a model that performs near the frontier, costs 40% less and can be customized or run in-house may be the more attractive option. Its very existence puts pressure on the pricing power of U.S. labs, the enormous valuations built around their technological edge, and the case for spending hundreds of billions of dollars on large data centers. While US might still have the edge currently, the rest of the world might choose a cheaper alternative & that should be a scary proposition for the US AI community.

Recent advances in artificial intelligence (AI) have raised hopes of a boost to economic growth. Many market participants believe that AI has the potential to be the most important general-purpose technology of our era. The recent inroads of generative AI in everyday applications in particular promise widespread efficiency gains. But we see a different picture. AI boosts GDP by raising productivity and investment, with variable effects on the demand and supply sides of the economy. The output gap is the difference between aggregate demand (GDP) and aggregate supply (potential output). At first, supply exceeds demand in most regions, due to the concentration of early adoption of generative AI (GenAI) in “easy-to-learn” tasks. This leads to a swift increase in productivity that significantly surpasses the initial investment. Subsequently, as AI investment extends to less profitable areas of “hard-to-learn” tasks, the situation evolves with demand outpacing supply. This results in a positive output gap that puts upward pressure on prices. In the widespread AI adoption scenario, the effects on the output gap are twice as large as in a conservative scenario. In short term, AI raises productivity and lowers marginal costs, exerting downward pressure on inflation. But in long term, by nature of AI, intensely dependent on energy & commodity space along with limits on learning skills, AI leads to an uptick in inflation. In a conservative scenario, commodity prices may rise by between 1% and 2% by 2033 due to an AI induced increase in demand. In the widespread adoption scenario, the projected effects are twice as great. A natural starting point in this analysis is the familiar view of AI as a productivity-enhancing general-purpose technology: by improving efficiency, compressing unit costs, and expanding effective capacity, AI can act as a structural disinflationary force. At the same time, the macroeconomic footprint of AI is not confined to productivity. Scaling and deploying AI requires costly complementary inputs and infrastructure, can reshape product-market pricing and market power, and can alter labor-market rents and wage-setting wedges. Moreover, the diffusion of AI has increasingly taken the form of an investment and infrastructure wave—most visibly through rapid expansion of compute capacity and data-center build-out—that operates as a demand impulse as well as a supply-side transformation. The sign, magnitude, and persistence of AI’s effect on inflation therefore cannot be inferred from a single channel or a single dataset; they are joint empirical and quantitative objects that depend on how these mechanisms interact in general equilibrium. The net inflation effect of AI is not mechanically disinflationary even when productivity gains are present. A rise in productivity is a mechanism that exerts a systematic downward force on inflation by reducing unit costs, but it competes with forces that can raise marginal costs or desired markups during diffusion. AI specific input costs arise because the deployment of AI at scale requires scarce complementary resources—compute, energy, cooling, networking, and specialized capital—whose shadow prices can rise when capacity expands rapidly or adjustment is slow. On the employment side, again we do not see any wide spread impact on employment due to AI adoption. Total job postings have been broadly stable, there is limited evidence of a change in the low firing environment across key layoff indicators & WARN indicators show no broad-based impact of AI as layoff remains low. To summarise, AI is still not a threat to US employment in general. In fact, wage growth remains the highest for graduates with bachelor degrees or higher degrees. There is limited relationship between wage growth & AI adoption according to average hourly earnings. So, when the new Fed Chair Kevin Warsh says that he believes AI will lead to marked productivity gains & lower inflation, he is only looking at short term trend. Long term AI is inflationary.

While geopolitical uncertainty has surged since the start of the US-Iran war in late February, unlike last year’s “Liberation Day” shock though, the war has had little impact on a nascent cyclical acceleration. The investment acceleration we had expected to take hold in 2026 arrived ahead of schedule, with 2025 marking the strongest full year of equipment capex growth since the post-GFC recovery. What began as a narrow surge in the tech/AI sectors has broadened to other industries and types of equipment spending. Capex has remained resilient in March and April. Capital goods shipments and imports have continued to accelerate along with domestic production of business equipment. Hyperscalers ramped up their capex guidance for FY 2026 in Q4 earnings calls around $130bn higher than what they forecasted in Q3. This was further pushed up by ~$55bn during Q1 earning announcements, partly due to higher prices amid capacity constraints. Despite the recent rise in interest rates, broader financial conditions appear accommodative, and bank C&I lending has continued to rise at a 15-20% annualized pace in recent weeks. Adding to the tailwinds for capex, tariff refunds have begun to ramp up. A cashflow windfall will likely support spending alongside strong growth in revenue and profits. The only weak spot seems to be consumer spending. Consumption has been resilient since the start of the US-Iran war. This strength is likely unsustainable though, with households relying on one-time stimulus cashflows to offset higher energy costs. We estimate households received ~$45bn in additional tax refunds this year due to the One Big Beautiful Bill Act passed last summer. For comparison, through May, we believe higher gasoline prices will have cost consumers ~$30bn. To summarise, we expect business investment to grow 7.8% in 2026 on a Q4/Q4 basis, driven by strong AI investment demand, expanded expensing provisions, and fading drags from the normalization of factory construction and tariffs. Coupled with last week’s GDP tracking data, this implies 2026 GDP growth of 2.1% on both a Q4/Q4 and full-year basis. It's on the inflation side where we see FOMC members turning hawkish last few weeks. Most notable was the hawkish pivot in an outlook speech by Governor Waller, who has represented the dovish core of the Committee. Waller indicated that his risk assessment had shifted towards inflation, which has replaced the labor market as the “driving force” behind monetary policy in the months ahead. Even household interest rate expectations are building in inflation expectations similar to financial markets. Specifically, the net interest rate expectations indicator from the Conference Board clearly shifted in the direction of foreseeing higher rates over the next twelve months. The current difference between the 2-year Treasury yield and the effective fed funds rate is ~40bps but this week high was 50 bps. Historically, that spread has been a reasonably good leading indicator for future changes in the fed funds rate. For example, over the past three decades, the difference between the 2y yield and fed funds rate leads the year-over-year change in the latter by roughly eleven months with a peak correlation of +66%. To summarise, recent Fed communications are consistent with a Fed that is well positioned near neutral but increasingly concerned that inflation may prove more persistent. While our baseline remains that the Fed is on hold near neutral indefinitely, we now see equal risks of a rate increase compared to a rate cut as the monetary policy outlook continues to be impacted by developments in the Middle East. We still think it is too early to have a convincing view on either a cut or a hike at least till we get clarity on the middle east conflict. We will be soon releasing a detailed piece on AI’s impact on US inflation. We believe while AI’s productivity gains put a cap on long term inflation, short term price impact of AI capex is leading to elevated goods inflation via semiconductor chips prices, computer flash memory & old computers.