Calamos Phineus Long/Short Fund Q3 2025 Commentary
tanit boonruen/iStock via Getty Images Key Points AI spending has reached mania levels, ignoring the core limitations of LLMs—their statistical construct, inability to reason sequentially, and commodity-like economics. That said, this cycle may end through gradual exhaustion rather than dramatic collapse. Trump’s multifaceted reflation agenda should support the broader non-AI economy into 2026. The ending of Fed intransigence, rising policy focus on housing, and planned bank capital overhaul reinforce the narrative of US economic resilience. Industrials remain the fund’s largest overweight. The data center buildout extends beyond GPUs—there is more to this infrastructure than semiconductor hardware. Cyclical positioning captures AI spillover effects without direct exposure to unsustainable capex dynamics at the ecosystem’s core. The challenge requires discriminating across divergent narratives rather than binary market timing. Balance and diversification are critical because the AI narrative has synchronized investor behavior to an unusual degree, amplifying reversion risk. The Return of Americo-Centrism The Calamos Phineus Long/Short Fund (CPLIX) rose 1.29% in Q3 (Class I shares at NAV), compared with gains of 7.8% and 4.4% for the S&P 500 Index and its equal-weighted brethren, respectively. Net exposure averaged 22% through Q3, down modestly from 24% at the end of Q2. For YTD 2025, fund NAV rose 11.36% with an average net delta-adjusted exposure of 21% (8.3% of alpha and 3.1% of beta return). Fund positioning reflected growing conviction that AI infrastructure narratives were approaching their apex, prompting gradual reduction of momentum-driven mega-caps while increasing exposure to pro-growth cyclicals. The long portfolio generated solid returns led by the industrials overweight which benefited from the spillover effects of the data center buildout. Short positioning proved challenging as momentum overwhelmed fundamentals. For global investors, this race to develop AI has rejuvenated the economic and financial dominance of the US.1 While apprehension about trade wars and stagflation has diminished since spring, a new anxiety has emerged: AI bubble risk. Our view: AI will either create more losers than winners among today’s technology leaders, or the incremental revenue opportunity is mightily exaggerated. While a souring in investor mood is certain, timing that inflection is the challenge. The equity-funded nature of today’s AI buildout—combined with hyperscaler monopolistic cash flows—may prove more resilient to traditional cost-of-capital limitations than historical technology booms. While “bubble” remains the dominant metaphor, others are possible including “death by a thousand cuts” or a gradual unwind ahead. This argues for balanced exposure: reducing concentration risk in AI infrastructure while capturing opportunities in cyclicals benefiting from Trump’s multifaceted reflation agenda. Back to the Future The US equity market’s recovery from April lows has been exceptional by any standard. Two explanations dominate: the first is one of US economic resilience led by AI-driven mega-caps, where the profitability divergence between the Magnificent Seven and the broader corporate world has defined this decade. Here, the central challenge remains timing the peak of investor euphoria around this super-cycle. The second explanation concerns the revisionism of Donald Trump. Throughout spring, most assumed his agenda would bring disruption. These fears were not mistaken, but the emphasis was misplaced: Trump 2.0 represents multifaceted reflation2 rather than dislocation for the US economy. We emphasize this distinction because Trump’s polarizing presence creates a polemical dust storm that obscures. Entering Q3, we anticipated late summer equity weakness would create opportunities ahead of an S&P 500 rally toward 7,000 in Q1 2026. With investors seeking pullbacks to buy, this has proven consensus thinking. Meanwhile, the crosscurrents multiply. Financial liquidity appears abundant while economic liquidity feels tight. The Federal Reserve is “behind the curve” and an incipient growth scare has emerged.
EQUITY MARKET PERFORMANCE Q3 2025 YTD S&P 500 7.8% 13.7% S&P 500 Equal Weight 4.4% 8.3% MSCI World ex-USA 4.8% 22.6% NASDAQ Comp 11.2% 17.3% Russell 2000 12.0% 9.3% MSCI EM 10.1% 25.2% MSCI Europe 3.2% 9.8% UK (FTSE 100) 6.7% 14.4% Germany (DAX) -0.1% 19.9% France (CAC 40) 3.0% 7.0% Australia (ASX) 3.6% 8.5% Italy (FTSE MIB) 7.4% 25.0% China (Shanghai Stock Exchange) 12.7% 15.8% Hong Kong (Hang Seng) 11.6% 33.9% Mexico (IPC) 9.5% 27.1% India (SENSEX) -4.0% 2.7% Brazil (Bovespa) 5.3% 21.6%
Click to enlarge
Data through 9/30/25. Note: Returns are price only in local currency. Past performance is no guarantee of future results. Source: Bloomberg. Indexes are unmanaged, do not include fees or expenses, and are unavailable for direct investment. Please see “Index Definitions” for additional information. Portfolios are managed according to their respective strategies, which may differ significantly in terms of security holdings, industry weightings, and asset allocation from those of the benchmark(s). Portfolio performance, characteristics and volatility may differ from the benchmark(s) shown.
Click to enlarge
AI Everything The obvious concern starts with the extraordinary over-investment and over-enthusiasm surrounding the AI narrative. The scale of activity3 and profitability sets it apart from the dot-com boom. Hyperscaler capex is running at an annual rate of $400 billion and is forecast to reach a cumulative total north of $3 trillion by decade’s end—equivalent to 10% of US GDP. To justify these outlays, annual sales exceeding $1.3 trillion in 2030 must emerge from somewhere.4 Consider the scale: OpenAI generated ~$4 billion in H1 2025. Yet Oracle (ORCL)’s >$317 billion spike in RPO backlog5 in the latest quarter—driven by OpenAI6 multi-year commitments—equates to 70% of the quarterly increase in US GDP. These companies are sinking billions of dollars into AI because AI capabilities could erode the moats surrounding their extremely profitable, near-monopolistic core businesses. This monopoly character of today’s technology leaders explains why all of this has evolved into a “super narrative.” Spending has progressed to “whatever it takes” levels that are unworkable for businesses that require a visible path to ROIs. This imperative to protect their moats is relentless—until the incremental value of the next LLM upgrade no longer extends or fortifies their competitive position. This treadmill is why the cost and capability progression of GPT models7 is consequential. ChatGPT-3 cost US $50 million to develop and launched in November 2022; ChatGPT-4 cost US $500 million. ChatGPT-5, using 80X more compute and budgeted at US $5 billion8 was released in August—it is a relative dud. Improved in some areas, regressing in others, the technology may be hitting a scaling wall of diminishing returns. Elon Musk speaks of 10x increases in compute driving 2x improvements in performance, yet the latest versions have fallen off this trajectory. Equally, this tendency to celebrate accuracy improvements on LLM benchmarks obscures a critical point: reliability rates must be consistently above 90% or higher. The jump from “sometimes works” to “dependably works” is where commercial value begins to emerge. This limitation is inherent in their design: LLMs are statistical simulations of human text. The finite complexity of language itself constrains how much these models can improve. Their statistical construct9 is why LLMs struggle to play chess despite “knowing” everything written about chess. They struggle with the sequential logic inherent in solving problems like counting tasks, basic algebra, map labelling and similar reasoning chains. Amidst the hyperbole around worker replacement, LLMs cannot manipulate representations of space, time, numbers, causality and sequence, or physical objects. Without these, they are mimicry tools with peripheral application. By peripheral, we refer to the divide between functionality and value: asking the right question versus any question.10 What are LLMs good at? By generating statistically likely generic responses, they excel at commonplace tasks like homework11 and basic software code. They falter when required to generate novel or complex outcomes. Rather than producing new or monetizable software solutions, LLM-assisted coding primarily reduces costs and leads to reduced commerciality across traditional software applications . LLMs will be important personal productivity tools. Yet retrieving and synthesizing existing knowledge differs from reasoning and creation. LLMs are like instant access to the right encyclopedia entry —a valuable capability, but one with constrained commercial potential. Recombining codified knowledge is inherently difficult to monetize at scale. Another challenge: unlike most digital businesses, LLM expenses do not decline with scale. Netflix (NFLX) exemplifies these scale economies: the incremental cost of streaming content to an additional subscriber is near zero. In contrast, LLMs operate inversely: more training and inference loads involve substantial incremental costs in the form of higher GPU usage.12 With their generic non-proprietary output, competitors can match incremental LLM improvements within months and with less compute power.13 The result is extreme price elasticity: any LLM that prices its service to recover the development or operating costs will lose share to lower-cost followers. These businesses are less software-like and more like capital-intensive commodity producers—think utilities. Many hope that the “human like” intelligence of LLMs can overcome these limitations with a breakthrough commercial application. Yet the nature of LLM output creates the same problem as for the calculator: all answers are non-proprietary. Contrast this to Netflix, which has a ~$500 billion market capitalization built on ~$ 45 billion in subscriptions with 300 million households because its content can’t be copied . Rather than viewing LLMs as the second coming of the internet, another analogy may be OpenSource 2.0. Linux and similar initiatives put advances in code in the public domain (much as AI coding promises to do), yet failed to generate valuable stand-alone businesses. The largest opensource business Red Hat was sold to IBM (IBM) for $32 billion 20 years later, a far cry from OpenAI’s most recent $500 billion valuation. The Capex Endgame History provides an instructive parallel. Today’s AI buildout echoes the capex boom of the late-1990s era. The scaling of internet and mobile telephony was central to sustaining “blue sky” expectations and associated valuations, but the emergence of price-elastic capacity ended the cycle. The pattern recurs across general-purpose technologies—from railways and electricity to radio, semiconductors, and the internet. These eras didn’t end because the dream fell short; they burst when the cost of capital began to rise. And the triggers for higher capital costs can be distinct from tighter monetary policy, including regulation (remedies ruling in Google’s antitrust case), increased competition (DeepSeek and open-source approaches), or buyers becoming more discriminating. Outside NVIDIA, the AI ecosystem14 is materially unprofitable. OpenAI will lose an estimated US $14 billion in 2025 assuming losses do not accelerate with more use of ChatGPT5. Anthropic faces the same challenge—losing more than 100% of revenue. The unusual intervention of NVIDIA, investing billions in equity directly into its largest customers signals that funding at the required scale is becoming harder. Assumptions around future profitability are necessarily sensitive to the inputs: Blackwell chip costs, electricity, financing, and GPU rental prices. Notably, NVIDIA extracts enormous rent from the industry;15 its unprecedented 80% gross margin (versus a more typical 50%) exacerbates the funding challenge. With its stunning returns, NVIDIA is strongly incentivized to keep the ecosystem afloat. So far, the industry has been creative in overcoming its financial challenges. GPU purchases are depreciated over six years versus a sub-two-year release cycle at NVIDIA.16 H100 rental prices have declined from $8/hour in Q2 2023 to less than $2/hour in 2025, yet the neo-clouds have raised debt backed by GPUs as collateral. Hyperscalers offer cloud services in exchange for equity in the LLM providers. This circularity of AI funding, with vendors like NVIDIA investing in customers and customers receiving warrants in suppliers echoes the cross-shareholding of Japan’s “keiretsu” model, which encourages industry cooperation. This is critical given the complexity of today’s GPU. The vulnerability: win-win arrangements become lose-lose when things go wrong. Sam Altman’s success is now everybody’s business; one cancelled project can become everybody’s problem. For now, today’s spending enjoys a positive feedback loop between rising investment and rising profits. AI infrastructure sellers like NVIDIA recognize profits immediately, while buyers like the hyperscalers depreciate such spending over years rather than up front. This dynamic works because of the monopoly-like characteristics that underpin cash flows—service quality and balance sheets can be degraded with little consequence.17 Until the “losers of AI” emerge, the spending party continues.18 OpenAI’s internal projections point to revenues of $200 billion in 2030, with profitability materializing dramatically as margins inflect from -25% in 2028 to +25% in 2030. Such trajectory is conceivable in a low-variable cost model, but rarely one as capital intensive as AI-related applications appear to be.19 All of this reveals the underbelly of today’s ‘strategic spending war’. The justification is that AI “will do everyone’s work,” and some have announced AI-related job cuts. But this is more euphemism than reality—obfuscating the rising financial pressure of building this infrastructure.20 As importantly, the moats of Meta (META) and Alphabet (GOOGL) are funded by cyclical advertising revenues now under threat. This interpretation explains why Apple (AAPL) has neither built nor invested in an LLM. Its monopoly rests on its App Store and iOS, and it doesn’t need an LLM to divert traffic: consumers are locked onto their platform with the purchase of an iPhone. Apple feels less vulnerable to LLM disruption than Amazon, Google or Meta. For similar reasons, Microsoft can be more disciplined across the AI narrative. Our conclusion: AI will either create more losers than winners among today’s technology leaders, or the incremental revenue opportunity is mightily exaggerated. Both point to an inevitable slowdown of spending. While a souring in investor mood is certain, timing that inflection is the challenge.21 Until that reckoning arrives, investors will run with the herd. The pattern is familiar: banks did so ahead of the Global Financial Crisis; the European telecoms did so in the late 1990s when they catastrophically overbid for 3G licenses. Recognizing a mania doesn’t immunize from its continuation—momentum by definition overwhelms fundamentals until it doesn’t. But recognition serves as a cautionary signal for portfolio construction, particularly regarding concentration risk in AI infrastructure beneficiaries. This vulnerability is broadly understood and uncontroversial. Yet investors must still grapple with the appropriate framework for their risk-adjusted positioning amidst these ahistorical narratives. The Silver Lining Paradoxically, excess capex is central to the adoption of transformative technologies. Euphoria lowers the cost of capital, allowing a more rapid buildout than normal economics might support. As the most valuable company on the planet with its low cost of capital, NVIDIA is incentivized to keep the party going. Its rapid product cadence is a prisoner’s dilemma for the hyperscalers, amplifying the cycle. When bubbles burst, excess capacity doesn’t disappear—it can be acquired at low prices by new or stronger players. This Schumpeterian destruction is inherent to technology manias and capitalism itself, giving more access to new capacity at lower prices than if the boom persisted . This outcome ensures the technology becomes embedded across society—but after the destruction unfolds. The good news is that the AI build-out has so far been a largely equity funded phenomenon, which differs from the capex booms that have more negative consequences because they are debt financed. Unlike the 2008 crisis, the policy and economic response could follow the 2000 playbook: equity benchmarks suffer, but damage to broader activity is less deleterious because interest rates fall and support domestic consumption. Today’s ‘bubble’ metaphor therefore requires care. Bubbles conjure a “burst,” implying that timing is paramount and a day late costs dearly. Yet much of today’s spending is funded by the hyperscalers’ enormous cash flows—financing less sensitive to traditional cost-of-capital shocks. Even as sentiment shifts or climaxes, these spending commitments can sustain momentum for the economy in the coming year. This argues for a different metaphor, one where deterioration unfolds through time or duration rather than sudden dislocation: “death by a thousand cuts”. AI equities could drift over quarters or years, exhaustion punctuated by failed rallies that attract fewer believers each time. This was how the bear market in crypto unfolded in 2021/2022. In sum, the absence of leverage implies outcomes that are “less bad” for Main Street, which is why Trump’s agenda may prove decisive. Without predicting when the boom ends, or whether it ends badly, history implies that as the economic significance of AI becomes clearer, the valuations of AI-linked stocks should fall. That is our assumption for 2026. Trump 2.0: The Agitation That Never Rests This narrative of rising risk for the dominant part of the equity world is conjoined with a second narrative: multifaceted reflation for the broader economy after years of profit recession. Here, investors must filter a deluge of over-politicized commentary. Apprehension about the stagflation consequences of Trump’s agenda has so far proven exaggerated. Segments of the US economy are under stress, yet macroeconomic figures have unfolded largely according to trend. Meanwhile, Trump has removed the roadblock of Fed intransigence. The recommencement of easing by the Federal Reserve and the administration’s focus on housing argues that the broader non-AI economy has support into 2026. While AI has underpinned the resilience of 2025, it has equally given rise to a “two-speed”22 economy. Fortuitously, this is beneficial for US equity assets: it allows the presumption of monetary reflation even in the context of solid consumer incomes and stubborn inflation. Equally important, the inflation math of higher tariffs has been mitigated by high elasticities of demand, substitution effects, and US dollar resilience. In the absence of a sudden end to the AI impulse, spillover into the non-technology areas points to revived leadership by cyclical parts of the market. Put simply, there is more to a data center business than a GPU, including multiple cyclical opportunities that can help counter mega-cap concentration risk. The challenge lies in managing exposure to both narratives while discriminating across their differing risk profiles. A dynamic liquidity setting dominates the backdrop. The health of private sector balance sheets is manifest, while innovation across the fintech world (including stablecoins, digital currencies, and crypto) rhymes with past episodes of non-bank credit creation. These innovations have historically been stimulative for the broader economy, reinforcing investor enthusiasm for risk-taking. And more “agitation”: the Trump administration is planning the biggest overhaul of US bank capital rules since the Global Financial Crisis. This removal of leverage constraints enables banks to finance the US deficit and economic growth. As another form of monetary easing, a recessionary contraction in credit is harder to envision, again supporting the outlook. Markets are enjoying this confluence of positive narratives. Yet much is dependent upon moderating economic activity—sluggish enough to support Fed rate cuts, but not so much as to strain corporate earnings. This narrow path would normally prompt a wider debate amongst investors, yet the AI narrative is synchronizing behavior to an unusual degree. This highlights the unusual character of today’s investment setting. Markets are “efficient” when they aggregate diverse viewpoints (long versus short horizons, fundamental versus technical approaches, growth versus value, etc.). In 2025, diversity has been overwhelmed by sentiment around the AI narrative—all visible in the outsized influence of the momentum style. When investors engage in too much ‘imitation’ and sentiment becomes uniform, this diversity is lost and markets become vulnerable to episodic, large moves. This argues for balanced portfolio positioning: increasing wariness toward AI-driven momentum stocks, but healthy exposure to pro-growth cyclicals on the assumption of sustained US expansion. Summary The Momentum of Americo-Centrism 2025 will be remembered as the year in which the AI narrative became an investment boom, if not a mania. Only in America could such substantive amounts of private capital be mobilized so rapidly in support of innovation. As importantly, it has been supported by US policymakers if only because this nascent industry is crucial to the Sino-American arms race. For global investors, this race to develop AI has rejuvenated the economic and financial dominance of the US. 23 While apprehension about trade wars and stagflation has diminished since spring, a new anxiety has emerged: AI bubble risk. Here, referencing the dot-com mania of the late 1990s is irresistible because many grasp the inevitable misallocation of capital. Less clear is what comes next. The benign view is that bubble anxiety is just the latest avatar of the wall of worry accompanying all bull markets. Indeed, this is how technological progress has always unfolded in a capitalist system. The conundrum is that the US market has become the victim of its own success: so internationally dominant and so comparatively expensive that it inevitably generates some form of anxiety. Many seek alternatives to this dominance and to the US dollar, and yet nothing compares with the growth-oriented culture of America and its reliable backstop of demand.24 Where pockets of overseas outperformance have emerged, these are cyclical value situations that are another derivative of US reflation. Their economies are too precarious to sustain investor confidence in the absence of benign American influence. The missing attribute for those economies that would challenge American leadership is the US consumer. Who would be the buyer of global production, if not America? Protectionist India? Impoverished Russia? Or a China whose control obsession stymies the emergence of a prosperous, independent middle class? And yet, a good part of the investment industry reasons as if we are still in a globalizing world. Debasement has become a popular and indiscriminate cry across the financial world. No Western government in any major economy is willing to arrest the rise of sovereign indebtedness.25 High levels of public debt, fiscal dominance, and the retreat of globalization would normally give rise to an inflation bias. This return of a term premium26 across equities and bonds feels inevitable. It is a risk for 2026. Investors feel they have little choice but to embrace this Americo-Centrism. For now, the “two speed” US economy is beneficial because it gives rise to monetary reflation as does the exit of Chair Powell in spring. Yet the power of the AI narrative blinds investors to the downside tails. We advise balance and diversification across client portfolios. Michael Grant, Co-CIO Fund Exposures and Attribution After the remarkable recovery from the April lows, we felt equities entered late summer from a position of vulnerability. Equity exposures have therefore been contained, averaging 21% net through the quarter comprised of an average gross long and short exposure of 83% and -62%, respectively. This is modestly below the 28% average since inception in 2002. Investors appear increasingly complacent about the evolution of AI, and the degree to which an increase in systemic risk could eventually start weighing on valuations. Historical parallels are numerous—including the dot-com bubble and housing boom—and serve to remind that the new economy companies will need to grow materially to overcome more normalized discount rates. Without presuming how the spending boom ends, or whether it ends badly, history argues that as the economic significance of AI becomes clearer, the valuations of AI-linked stocks should compress. In other words, as investors become more certain about how big this becomes, the same systemic risk as that embedded in the “old economy” raises the discount rate. Stock selection within the long portfolio was sufficiently strong to offset the headwind of underweighting the Mega-caps. Long positions contributed 6.4% in Q3, nearly matching the 6.6% return that a similar S&P 500 exposure would generate. The short book detracted -5.1%, driven almost entirely by index hedges and, to a lesser extent, Apple (-29 bps). The fund increased long exposure to traditional cyclicals including industrials (+4.0%), consumer discretionary (+2.6%), and materials (+2.3%) while reducing technology (-2.7%), healthcare (-1.3%), and consumer staples (-0.8%). Through the July release season, stock reactions often reflected positioning excesses rather than fundamental surprises. Calamos Phineus Long/Short Fund (CPLIX) Returns by Strategy
STRATEGY CONTRIBUTION TO PERFORMANCE Q3 YTD Long 6.38% 19.02% Short -5.09% -7.66% Net 1.29% 11.36%
Click to enlarge
Q3 Returns by Sector (Long Strategy)
SECTOR CONTRIBUTION TO LONG BOOK Information Technology 2.65% Industrials 1.54% Communication Services 1.09% Health Care 0.64% Financials 0.49% Materials 0.15% Energy 0.06% Consumer Discretionary -0.08% Consumer Staples -0.16% Total 6.38%
Click to enlarge
Average Delta Adjusted Net Exposure by Sector
SECTOR AVERAGE FOR Q3 Industrials 31.46% Information Technology 13.77% Consumer Discretionary 8.45% Financials 6.91% Health Care 6.67% Communication Services 5.81% Materials 3.08% Consumer Staples 1.25% Energy 0.07% Index Hedges -59.05% Total 18.42%
Click to enlarge
Footnotes 1 The AI boom is the exemplary expression of American exceptionalism. Despite its international dominance and comparatively expensive valuation, America has no apparent replacement for its place at the center of the modern monetary order. The underpinning of this is its ability to provide both stability and demand for the world economy as needs require. 2 Unrelated but coordinated policy efforts to stimulate demand and incomes across the US private sector. 3 By some estimates, the buildout of data centers led by NVIDIA (NVDA) GPUs has contributed more than 1% to GDP growth in 2025; one can only guess at the indirect wealth impact of AI-related share gains upon GDP. 4 $3 trillion in capex depreciated over six years is $500 billion annually. Adding $180 billion in financing expenses at 6% and power and other costs of approximately $100 billion (75GW at $0.15 per kWh) brings the total cost to $780 billion for the cloud providers. A conservative margin of 30% implies an annual price tag for purchasers (LLM services of firms like OpenAI and Anthropic) exceeding $1 trillion. OpenAI internally projects a 70% gross margin in 2029, which implies revenues much higher than this. 5 This unprecedented spike in Oracle’s Remaining Performance Obligations (RPO) backlog is the total value of contracted revenue a company hasn’t yet recognized but is contracted to deliver. 6 In the three months prior to the AMD deal, OpenAI has committed close to $500 billion of spend phased over multiple years ($300 billion for cloud computing with Oracle, $100 billion with NVIDIA, HBM memory of circa. $70 billion, $10.5 billion additional spend at Coreweave (CRWV), $10 billion at Broadcom (AVGO) for AI accelerators). 7 GPT-5 was released on August 7; a Polymarket poll showed OpenAI falling from 75% to 14% in the space of an hour on the question of which company would have the best AI model. OpenAI had to restore access to previous models for paid users due to the backlash. 8 GPT-5 was originally scheduled for Q3 2023, but it was delayed for additional training and videos and released as 4.5 in early 2025; after additional interface work, 5.0 was released in August. These cost estimates are consistent with OpenAI’s reported revenue and expense lines for 2024/25. 9 The weakness of probabilistic language generators becomes more apparent in images and videos, where they are harder to generate with vector correlation and hallucinations are clearly visible. 10 Determining the right question or answer requires human experience with reality; it represents the difference between perception versus conceptualization. LLMs struggle to “know” which tool or rule to apply for any work sequence, which naturally constrains utility. 11 According to OpenRouter data, ChatGPT usage peaked on May 27 at 97.4 billion tokens per day during finals season compared to 36.7 billion tokens per day in June when schools let out. 12 Current costs vary widely from $2-$10/GPU hour in community clouds to $12/H100 hour for on-demand access to major cloud providers. 13 Competitors can learn from publicly available results rather than repeating the full R&D cycle. DeepSeek showed that inference costs could be reduced by 90% with this optimization. 14 Includes the independent AI data centers such as Coreweave, Lambda, Nebius (NBIS) and Hyperbolic. For the hyperscalers Amazon (AMZN), Google and Microsoft (MSFT), margins on AI-related workloads are well below those of traditional workloads. By some estimates, Microsoft Azure margins for AI were breakeven in Q4 of 2024 versus >70% for non-AI workloads. 15 NVIDIA earns gross margins north of 70%—previously unheard of for a hardware company. Its evolution from a chip supplier to a vertically integrated supplier of rack-scale server and networking systems to complete data center architectures has squeezed out everyone else. 16 The intent is to continue running the older technology because the cost is sunk. Yet, it would be cheaper to replace the older generation with the latest GPUs rather than implement the latest GPUs in an entirely new data center. 17 In her antitrust ruling, Judge Brinkema highlighted the declining quality of Google search, with ads and promoted content appearing first as a signal of its monopoly power. 18 Total spending for the Mag 7 across 2024 and 2025 looks to be US $560 billion versus ~$35 billion in revenues. 19 Meta currently generates ~$200 billion of revenues with 80% gross margins, valued at $1.8 trillion. OpenAI is projecting it can deliver $200 billion in revenues at 70% gross margins by 2030, but the contrasting nature of their business models implies a far lower valuation is appropriate. 20 Based on “TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks” by Frank F. Xu. The paper presents a benchmark focused on real-world work tasks and concludes a significant gap exists for AI agents to autonomously perform most jobs that human workers would do, even in a relatively simplified benchmarking setting. 21 The most likely catalyst may be growing awareness of the limitations of the newer training models, which diminishes the imperative for an ever-escalating commitment. 22 New economy activities are booming; old economy sectors are stagnant. 23 The AI boom is the exemplary expression of American exceptionalism. Despite its international dominance and comparatively expensive valuation, America has no apparent replacement for its place at the center of the modern monetary order. The underpinning of this is its ability to provide both stability and demand for the world economy as needs require. 24 The ongoing competition with China creates considerations of strategic risk and the degree to which a fiduciary should place client monies in countries not aligned with the US. Even strategic non-alignment can be dangerous in a polarizing world as India has discovered. 25 The three largest European economies are in trouble, beset by economic stagnation, weak leadership and popular discontent. An authentic political or fiscal crisis is more likely to unfold in Europe or Japan than the US. 26 Term premium: the extra compensation locking up capital over time – for holding longer-term assets instead of rolling over short-term investments. About the Author Michael Grant,Co-CIO, Head of Long/Short Strategies, and Senior Co-Portfolio Manager Michael Grant manages investment team members and leads the portfolio management team responsible for our Long/Short strategies. He is also a member of the Calamos Investment Committee, which is charged with providing a top-down framework, maintaining oversight of risk and performance metrics, and evaluating investment processes. He joined Calamos in 2015 and has more than 35 years of investment industry experience. Prior to joining Calamos, Michael founded Phineus Partners in 2002, where he launched the Phineus long/short strategy. Previously, he was a Managing Director of Schroder Investment Management with responsibilities over US equity mandates. During his tenure at Schroders, he also served as Head of the Global Technology Team and Head of the US Equity Team in London. Prior to that, Michael was a portfolio manager for the National Investment Trust Co. in Taipei, Taiwan and a US equity analyst for the Principal Group in Canada. Michael earned a master’s degree from the London School of Economics, where he specialized in International History. He has Bachelor of Commerce from the University of Alberta, Canada. To learn more about the potential benefits of including Calamos Phineus Long/Short Fund (CPLIX) in an asset allocation, please contact your Calamos Investment Consultants at 866-363-9219. Past performance is no guarantee of future results.
Click to enlarge
Original Post Editor’s Note: The summary bullets for this article were chosen by Seeking Alpha editors.
已发布: 2025-12-01 06:30:00
来源: seekingalpha.com
