NeutralAMZNAmazonAmazonGOOGAlphabetAlphabetGOOGLAlphabetAlphabetMETAMSFTMicrosoftMicrosoftTBLL

1w ago on

US Midterms 2026: Democratic Control of Congress, Political Gridlock, and the Implications for Global Markets The Democrats’ projected victory in both the House and Senate represents a significant shift in the American political landscape, particularly because the party managed to capture Senate seats in traditionally Republican states such as Texas, Kansas and South Carolina, suggesting that voter dissatisfaction with inflation, living costs and the administration’s handling of the economy has extended well beyond the usual competitive districts. However, I think the most important implication for financial markets is not that Democrats have gained control of Congress, but that the United States is entering a period of divided government in which legislative gridlock, fiscal negotiations and regulatory scrutiny will become increasingly important, potentially constraining President Trump’s ability to advance new legislation while leaving much of his executive authority over tariffs, foreign policy and existing regulations intact. The new congressional balance takes effect in January 2027, and while the Democrats will gain substantial oversight and legislative powers, they will not automatically gain the ability to reverse existing policies because Trump retains his presidential veto and most Senate legislation remains subject to the 60-vote threshold. From a macroeconomic perspective, the most consequential development could be the return of fiscal policy as a major source of market uncertainty, particularly when the United States is already struggling with substantial budget deficits, elevated interest expenses and an increasingly difficult Treasury financing environment. Democratic control could constrain additional deficit-financed tax reductions or discretionary spending initiatives, which might eventually help moderate Treasury issuance relative to a more expansionary policy baseline, but I would be reluctant to assume that political gridlock automatically translates into fiscal discipline because neither party has demonstrated a particularly strong willingness to address the structural imbalance between government expenditure and revenue. More importantly, negotiations over government funding and the debt ceiling could become considerably more contentious in 2027, introducing risks of shutdowns, delayed fiscal agreements and additional volatility in Treasury markets. For investors, the distinction between reducing future deficits and merely preventing further fiscal expansion is enormous, particularly when the existing debt stock and associated interest burden remain substantial regardless of which party controls Congress. This brings us to the bond market, which I believe will remain a more important determinant of equity valuations than the election result itself, because even if divided government reduces expectations for additional fiscal stimulus, that does not automatically mean long-term Treasury yields will decline. We need to distinguish between weaker growth expectations, which could pull yields lower, and a higher fiscal or political risk premium, which could push longer-dated borrowing costs higher even when growth is moderating. A prolonged confrontation over the debt ceiling could also create substantial volatility in short-dated Treasury bills and government liquidity without necessarily producing the same reaction across the entire yield curve. The real question is whether the new political configuration improves confidence in America’s long-term fiscal trajectory or simply introduces another layer of uncertainty into an already fragile bond market, and I think that distinction matters enormously when equity valuations remain sensitive to the discount rate applied to future earnings. The implications for technology and AI are particularly interesting because a Democratic Congress would have substantially greater authority to investigate large technology companies, examine AI safety practices and scrutinize the environmental and infrastructure costs associated with the enormous data-center expansion taking place across the United States. This could introduce additional uncertainty for hyperscalers such as Microsoft, Amazon, Alphabet and Meta, particularly if congressional investigations increase pressure for stricter model evaluations, greater transparency, data-center permitting requirements or additional scrutiny of electricity and water consumption. The change in state governments also matters, particularly in Texas and Georgia, where data-center investment has become increasingly important and state-level permitting, electricity costs and infrastructure policy can influence project economics even without major new federal legislation. However, I would strongly distinguish regulatory friction from a deterioration in underlying AI demand, because the commercial incentives for enterprises to adopt increasingly capable AI systems remain intact, while additional permitting constraints could actually increase the value of existing power connections, operational data centers and secured infrastructure capacity. I remain structurally constructive on AI, although the election introduces another variable investors should incorporate into the timing and cost of future infrastructure deployments. Trade policy is another area where I would avoid overly optimistic conclusions because Democratic control of Congress does not automatically mean that Trump’s tariffs will disappear or that relations with China will normalize, particularly when considerable trade authority remains delegated to the executive branch and concerns surrounding advanced semiconductors, national security and technological competition have become increasingly bipartisan. Congress could increase scrutiny of trade measures, challenge particular policies or negotiate restrictions on future executive authority, but meaningful legislative changes would still face presidential opposition and existing legal constraints. Consequently, I would not interpret the election as an immediate catalyst for a substantial reversal in semiconductor export restrictions, a broad normalization of US-China trade or a reduction in geopolitical risk. Similar uncertainty applies to defense spending, healthcare, banking and cryptocurrencies, where congressional investigations and legislative bargaining could change the policy environment, although major statutory changes would require broader political agreement. For equities, I think the immediate reaction could be complicated because political gridlock can reduce the probability of major new legislation while simultaneously increasing uncertainty around budgets, regulation and executive policymaking, meaning that different sectors could experience very different outcomes. Highly regulated industries such as banking, pharmaceuticals and cryptocurrency-related businesses may face greater congressional scrutiny, while certain renewable-energy and healthcare initiatives could receive stronger legislative attention, although the ability to translate those priorities into law remains limited by the presidential veto. Technology companies may face additional headline volatility, particularly given the ongoing reassessment of AI capital expenditure and elevated valuations, but I would caution against assuming that a Democratic Congress necessarily undermines the earnings trajectories of companies benefiting from structural growth in computing, cloud infrastructure, semiconductors and automation. Ultimately, markets will continue differentiating businesses according to revenue growth, operating margins, cash generation, returns on invested capital and valuations, rather than permanently repricing entire industries based on which party controls Congress. My broader conclusion is that the Democratic sweep represents a meaningful political constraint on the Trump administration but does not fundamentally change the structural economic forces driving US markets, particularly the expansion of AI infrastructure, the extraordinary capital requirements associated with that investment cycle, persistent fiscal imbalances and the increasing sensitivity of equity valuations to long-term interest rates. The next two years could involve more congressional investigations, slower legislative progress, contentious fiscal negotiations and greater uncertainty surrounding sector-specific regulation, but those developments should be evaluated according to their actual impact on corporate earnings, capital expenditure and financial conditions rather than through a simple partisan framework. For investors, I would pay considerably more attention to the direction of Treasury yields, the sustainability of corporate earnings and the economics of AI infrastructure than to the political headlines themselves. Washington can influence market sentiment, but over the longer term, earnings, cash flow and the cost of capital will remain the variables that ultimately determine investment returns.

AMZNAmazonBullishAmazon

1w ago on

AMAZON $AMZN JUST DELIVERED ITS FASTEST AWS GROWTH IN FIVE YEARS. THE STOCK CRASHED 8% AFTER HOURS. Welcome to a market where even extraordinary AI growth is no longer enough. Look at Amazon's Q3'26 numbers. Revenue: $201.9B, +12% YoY Operating profit: $28.1B, +61% YoY AWS revenue: $46.0B, +39% YoY AWS operating profit: $17.8B Operating margin: 13.9%, a record high AWS is now running at a $184 BILLION annualized revenue rate, growing almost 40%. Absolutely extraordinary for a business of this scale. Management even expects AWS growth to exceed 40% in both 2027 and 2028. So why did the stock collapse? BECAUSE THE AI BOOM IS BECOMING EXTRAORDINARILY EXPENSIVE. Amazon spent $77.3 BILLION on gross capital expenditure in ONE QUARTER, approximately $20 billion above expectations. 2026 cash capex guidance: $220 BILLION. 2027 spending? Even higher. Meanwhile, trailing-twelve-month free cash flow has deteriorated to NEGATIVE $42.4 BILLION. That's the uncomfortable reality. Amazon is generating record operating profitability while simultaneously consuming enormous amounts of cash to finance its AI ambitions. And the market is beginning to demand something beyond impressive AI revenue growth. SHOW US THE RETURNS ON THAT CAPITAL. EPS also missed at $1.89 versus $1.96 expected, despite operating income beating consensus, with higher financing costs contributing to the pressure below the operating line. Fourth-quarter operating income guidance was also below expectations. At 22.9x FY27 forward earnings before the selloff, Amazon is reasonably valued relative to its long-term growth potential, but the enormous capex requirements cannot simply be ignored. Here's my investment view. WE OWN AMAZON. WE REMAIN LONG AMAZON. BUT WE ARE NOT ADDING YET. AWS is accelerating, enterprise AI demand remains exceptional, and Amazon's underlying operating businesses are performing remarkably well. However, given the ongoing correction across AI and technology equities, I prefer to wait for a more attractive entry point before increasing exposure. The AI revolution is real. The revenue is real. The profits are increasingly real. But so is the investment bill. And ultimately, shareholders are paid from cash flows, not capital expenditure announcements. LONG $AMZN. LONG AI.

AMZNAmazonBullishAmazonGOOGAlphabetBullishAlphabetGOOGLAlphabetBullishAlphabetNVDANVIDIABullishNVIDIATSMTSMCBullishTSMCINTCIntelBearishIntel

1w ago on

INTEL $INTC JUST CRASHED 24% AFTER HOURS DESPITE BEATING EARNINGS EXPECTATIONS BY 30%. BRUTAL. Welcome to a market where even a strong quarter means NOTHING if your forward guidance disappoints. Intel's Q3'26 results were actually impressive. Revenue: $16.5B, +21% YoY Adjusted EPS: $0.52 vs. $0.40 expected Gross margin: 44.1%, +410bps YoY Data Center & AI: $7.3B, +78% YoY But then came the guidance. Q4 revenue: $15.5B midpoint vs. $17.1B expected. Q4 adjusted EPS: $0.25 vs. $0.43 expected. Gross margin expected to collapse from 44.1% to 38%. And the market absolutely HATED it. Here's what makes Intel particularly interesting. THE AI BOOM IS HELPING INTEL'S SERVER BUSINESS WHILE HURTING ITS PC BUSINESS. Exploding AI infrastructure demand is driving server CPU sales, but the same boom is contributing to soaring memory prices, making PCs more expensive and depressing consumer demand. Meanwhile, Intel Foundry remains deeply unprofitable, losing approximately $1.85B in Q3, despite reporting $6.7B in revenue, most of which comes from internal manufacturing activity rather than external customers. And here is my biggest concern. INTEL WAS TRADING AT 42.2x FY27 FORWARD EARNINGS BEFORE THE SELLOFF. That's an extraordinary valuation for a company still navigating a difficult turnaround, substantial manufacturing risks, persistent foundry losses and significant capital requirements. Even after a 24% decline, the stock would mechanically trade around 32x forward earnings, assuming estimates remain unchanged. And remember, Intel recently completed a $23B equity offering, introducing substantial shareholder dilution. I believe Intel's operational recovery is genuine, particularly in server CPUs, but improving fundamentals do not automatically make a stock attractive when the valuation already assumes a successful turnaround. We currently do NOT own Intel, and the 24% collapse alone is not enough to make us buyers. I remain much more comfortable owning established AI infrastructure winners such as $NVDA, $TSM, $GOOG, $AMZN and $SKHY. The AI boom is real. But not every semiconductor company deserves an AI valuation. The market rewards earnings today, but ultimately prices the earnings of tomorrow.

NeutralAMZNAmazonAmazonGOOGLAlphabetAlphabetMSFTMicrosoftMicrosoftNVDANVIDIANVIDIA

2w ago on

Anthropic: Cheaper Tokens, Exploding Demand, and the $100 Billion ARR Question I am shocked, absolutely shocked, that Anthropic’s token consumption has gone vertical after the company substantially reduced inference prices, except that I am not really shocked at all, because this is precisely the dynamic I have been discussing for months, namely that the AI industry is increasingly demonstrating how falling unit prices do not necessarily translate into falling industry revenues or weaker demand for computing infrastructure, particularly when cheaper intelligence allows companies to deploy AI across a much broader range of workflows, and the latest OpenRouter data provide another compelling example of how making intelligence more affordable can dramatically accelerate the amount of it that customers are willing to consume. The numbers are extraordinary, with Anthropic’s daily token consumption on OpenRouter increasing approximately 3.6x since the launch of Claude Opus 5.5 on September 22, as the seven-day moving average rose from roughly 600 billion tokens per day to approximately 2.2 trillion by October 24, while individual daily observations approached 3 trillion tokens, representing a remarkable acceleration after months in which consumption had remained comparatively stable. What makes the chart particularly interesting is not only the magnitude of the increase but also the timing, because the acceleration began after the introduction of Opus 5.5, continued after Sonnet 5.5 arrived on September 28 and strengthened further following the October 7 launch of Haiku 5.5, suggesting that the combination of improved model capabilities, lower inference costs and a broader range of economically viable workloads is stimulating substantial incremental demand, rather than merely causing customers to substitute one Anthropic model for another. Anthropic has aggressively improved the economics of its model lineup, with Haiku 5.5 costing approximately 75% less to run than its predecessor on average, while Sonnet 5.5 offers materially better price-performance than Sonnet 5, and additional reductions in caching costs have made agentic workloads considerably cheaper, which matters because agents can generate far more inference activity than conventional chatbot interactions as they repeatedly reason, retrieve information, call tools, revise outputs and continue working through multi-step tasks. And the results are increasingly visible in consumption, because the lower cost of each individual interaction appears to be encouraging developers to use models more frequently, to assign them more complex tasks and to embed them more deeply into production workflows that previously would have been considered too expensive or too unreliable to automate. The second chart reinforces the same point from the pricing side, with Anthropic’s blended token price falling toward approximately US$0.80 per million tokens, compared with levels closer to US$1.50 only a month earlier, while OpenAI’s corresponding price index has also declined substantially, illustrating that the effective price of machine-generated intelligence is falling rapidly even as the amount of intelligence being consumed appears to be accelerating. Updated October 25 data put Anthropic’s blended token price around US$0.85 per million tokens, down approximately 43.5% over the preceding month, illustrating how quickly the effective cost of intelligence is declining as model providers cut prices, improve efficiency and customers shift between models, although the precise revenue implications still depend on the mix of models being used, the distribution channel through which customers access them and the extent to which lower prices stimulate additional workloads. This creates what I think is one of the most important economic questions in the entire AI investment cycle: What happens when the price of intelligence falls dramatically, but the quantity of intelligence consumed increases even faster? In traditional commodity markets, a significant increase in supply tends to pressure prices and eventually discourage additional investment, but intelligence is different because cheaper inference can expand the range of economically viable applications, allowing businesses to automate tasks that previously made little financial sense, while increasingly autonomous agents can consume substantial amounts of additional compute without requiring a proportional increase in human users, meaning that the market can grow through greater intensity of usage even before it grows through a comparable increase in the number of customers. Consider a company that previously spent US$100,000 per month on AI inference, perhaps using models only for a narrow set of customer-service, coding or analytical tasks because the cost of deploying them more broadly was difficult to justify. If the effective price of inference falls by 50%, the company could maintain exactly the same workload and reduce spending to US$50,000, which would obviously be negative for the model provider’s revenue if nothing else changed, because the provider would be delivering the same amount of intelligence for half the price. But that is only the static calculation, and static calculations are often the least useful way to think about technologies whose costs are falling rapidly and whose applications are still expanding. If the lower price allows the company to expand AI deployment across customer service, software development, financial analysis, procurement, logistics and internal operations, while also allowing employees to use more capable models for tasks that were previously handled manually, total consumption could increase fivefold or tenfold, meaning that the customer might ultimately spend substantially more despite paying much less for each individual unit of intelligence. That is the economic mechanism I find so compelling, because the relevant question is not simply whether the price of a token is falling, but whether the decline in price is unlocking enough additional usage to expand the total pool of revenue, compute demand and economic activity associated with AI. Lower prices expand the addressable market, which increases consumption, which creates demand for additional compute, which encourages further infrastructure investment. This resembles the Jevons paradox, where improvements in the efficiency or affordability of a resource can increase its aggregate consumption rather than reduce it, although the analogy should not be treated as a mechanical economic law because the outcome still depends on demand elasticity, application development, customer budgets, reliability, latency, integration costs and the many expenses associated with deploying AI beyond the inference itself. And the Anthropic data are particularly interesting because the increase in consumption does not appear to be driven exclusively by the cheapest model, which would make the story more ambiguous because a surge in low-cost tokens could reflect substitution without necessarily representing a comparable increase in revenue or economic value. A substantial portion of the incremental volume comes from Claude Opus 5.5, Anthropic’s premium model, which suggests that developers are not merely switching to the lowest-cost intelligence available but are also expanding their use of more capable models when the resulting productivity gains justify the cost, particularly in coding, reasoning and agentic workflows where a more capable model can complete a task with fewer errors, less supervision and fewer costly iterations. That distinction matters enormously, because if all the growth were coming from extremely cheap models replacing expensive models, token consumption could rise dramatically without necessarily producing equivalent revenue growth, and the apparent expansion in usage could conceal a deterioration in the provider’s realized price per unit of output. But if premium models are simultaneously gaining substantial usage, that suggests customers are increasingly willing to pay for more sophisticated reasoning, coding and agentic capabilities where better performance creates meaningful economic value, and it also suggests that the market is not simply becoming a race toward the lowest possible price but is instead becoming segmented according to the value that different levels of intelligence can create. And this brings us to Anthropic’s revenue, where the relationship is straightforward in principle but much more complicated in practice: Revenue growth depends on consumption growth multiplied by realized revenue per token. If token volume rises 3.6x while the effective price per token falls approximately 43%, the simplified revenue implication would still be roughly 2x the original revenue, assuming the same customer mix, billing structure and comparable measurement periods, because a 3.6-fold increase in volume multiplied by approximately 57% of the original price still produces a result slightly above twice the original revenue. That is an extraordinary result, because it means a company can nearly halve its effective pricing and still potentially double revenue if consumption responds strongly enough, which is precisely why falling prices should not automatically be interpreted as evidence that the market is becoming less attractive. However, this calculation is illustrative rather than an estimate of Anthropic’s actual revenue growth, because OpenRouter represents only one distribution channel, its model mix may differ materially from Anthropic’s broader customer base, the timing of usage and billing may not align perfectly, enterprise contracts may have different economics and the Ornn pricing index measures a different population of transactions. So I would not conclude that Anthropic’s October revenue has already doubled simply by multiplying the two charts, because the charts demonstrate a powerful demand and pricing dynamic but do not provide a complete view of the company’s consolidated revenue, gross margins or customer mix. But the direction is undeniably encouraging, particularly because the data suggest that Anthropic is benefiting from both sides of the equation at once, with lower prices encouraging more usage while improved model capabilities make customers more willing to deploy the technology in economically important workflows. And this is where the speculation around US$100 billion-plus in year-end annualized revenue becomes particularly interesting, because the question is no longer whether Anthropic has found product-market fit, but whether the company can convert its rapidly expanding usage into a revenue run rate that would place it among the fastest-growing large technology businesses in history. Anthropic reported US$16.3 billion of revenue during the first half of 2026, while reported third-quarter revenue of approximately US$16.8 billion would imply an annualized quarterly revenue run rate of roughly US$67 billion, assuming that figure is confirmed in the company’s subsequent financial disclosures and represents a comparable measure of revenue rather than bookings, contracted revenue or another operating metric. To reach a US$100 billion annualized revenue run rate by December, Anthropic would need to exit the year generating approximately US$8.3 billion per month, compared with the roughly US$5.6 billion monthly average implied by the reported third-quarter figure, which means that the company would need to increase its monthly revenue run rate by a substantial amount in a very short period. That requires an approximately 49% increase over the third-quarter average monthly revenue level, a demanding target for any business and an extraordinarily demanding target for a company already operating at this scale, because maintaining rapid percentage growth becomes increasingly difficult as the revenue base expands. But is it impossible? I certainly would not dismiss it, because the underlying usage data suggest that Anthropic may be operating in a market where lower prices, better models and expanding agentic workloads can reinforce one another, creating a growth curve that looks less like a conventional software ramp and more like the early stages of a new computing platform. If consumption across Anthropic’s broader ecosystem is accelerating anything like what OpenRouter is showing, while enterprise adoption continues expanding and customers increasingly deploy agents into production environments, then the possibility of a US$100 billion annualized exit rate becomes more credible than it would have appeared only a few months ago, particularly if direct enterprise contracts and cloud distribution channels are growing faster than the public API data alone would imply. Nevertheless, possible is not the same as proven, and the critical missing evidence is whether the acceleration in tokens is translating into a comparable increase in realized revenue across Anthropic’s direct API, enterprise subscriptions, AWS, Google Cloud, Azure and other distribution channels, because usage can grow rapidly while revenue conversion varies depending on discounts, committed-use agreements, reseller economics and the mix between premium and lower-cost models. That is the number I would want to see in the next S-1 amendment, along with clearer disclosure on revenue by channel, gross margins by model family, committed compute obligations, customer concentration and the amount of revenue that comes from recurring enterprise contracts rather than usage that may fluctuate with pricing or experimentation. The other side of this development is arguably even more important for the broader AI infrastructure trade, because the same data that suggest falling prices are stimulating demand also suggest that the physical resources required to produce intelligence remain scarce. GPU rental prices are strengthening even as token prices are falling. The latest Ornn market data show H100 rental pricing around US$3.17 per GPU-hour, with its index approximately 25% higher over the preceding month, while H200 and B200 rental benchmarks have increased by roughly 30% and 21%, respectively, creating a striking divergence between the falling price of AI output and the rising price of the computing capacity required to generate it. This is an extraordinary divergence, because customers are paying less for the intelligence produced by AI systems while the market is simultaneously assigning greater value to the computing resources required to produce that intelligence, which suggests that demand for inference and training capacity is expanding faster than supply can adjust. And this is precisely why I have consistently argued that token price deflation should not automatically be interpreted as evidence of weakening AI demand, because in a market with strong demand elasticity, lower prices can stimulate enough additional usage to increase the total amount of compute required even as the cost of each individual output declines. It can mean exactly the opposite, because lower token prices make more applications economical, those applications generate additional inference workloads, more inference increases the demand for accelerators and supporting infrastructure, and the resulting competition for scarce computing capacity can sustain or even increase GPU rental rates. The distinction is that intelligence is becoming cheaper at the application level while the physical capacity required to produce intelligence remains scarce, and that is a very different phenomenon from a conventional demand collapse in which falling prices reflect excess supply, weak utilization and declining willingness to pay. But there is an important qualification here, because rising GPU rental prices alongside falling token prices are not automatically positive for every company in the AI ecosystem, and the same market conditions that create strong demand for infrastructure can also create intense pressure on the companies that must purchase or lease that infrastructure. For infrastructure owners, higher rental rates can be beneficial when utilization remains strong and financing costs are controlled, because each additional hour of GPU availability can generate more revenue and the scarcity of capacity can improve pricing power, although the economics still depend on depreciation, power costs, networking, maintenance and the risk that a new generation of hardware makes existing equipment less competitive. For semiconductor manufacturers, sustained demand for additional compute supports equipment orders, capacity expansion and potentially attractive pricing, particularly when customers are competing to secure advanced packaging, high-bandwidth memory and leading-edge manufacturing capacity. For model developers, however, the relationship is more complicated because falling realized token prices combined with higher compute costs can compress gross margins unless improvements in model efficiency, batching, caching, hardware performance and utilization more than offset the pricing pressure, meaning that a model provider can experience explosive revenue growth while still facing difficult questions about the profitability of each additional unit of usage. This is why I would distinguish between the growth of the AI market and the distribution of profits within the AI market, because a larger market does not guarantee that every participant captures an attractive share of the value created. AI consumption can explode while individual model providers struggle to retain pricing power, particularly if customers can switch between competing models or if cloud platforms use their distribution advantages to negotiate lower prices. GPU demand can increase while certain infrastructure operators earn disappointing returns because they financed equipment at excessive valuations, overestimated utilization or underestimated the speed at which newer hardware would reduce the value of their existing fleets. And the semiconductor supply chain can enjoy strong orders even while the companies purchasing the hardware compete aggressively enough to reduce their own margins, because the strategic importance of AI can encourage businesses to spend heavily before they have fully demonstrated that the resulting workloads will generate attractive returns. These outcomes are not mutually exclusive, and in fact they may all occur simultaneously as the industry expands. That is also why Anthropic’s IPO prospectus becomes even more interesting when viewed alongside these charts, because the company has demonstrated extraordinary commercial growth, improving operating leverage and an expanding base of large enterprise customers, but it has also committed enormous amounts of capital to securing computing capacity for years into the future. If intelligence consumption continues expanding at the current pace, those commitments could prove extraordinarily valuable because Anthropic is effectively securing production capacity ahead of a potentially enormous demand curve, and the company may benefit from having access to compute that competitors or new entrants cannot obtain as easily during periods of scarcity. But if token prices continue collapsing while the cost of compute remains elevated, the company will need to become dramatically more efficient at converting GPU-hours into commercially valuable output, because revenue growth alone will not be enough if each additional dollar of revenue requires an even larger increase in infrastructure spending. The critical metric is not simply tokens consumed, but gross profit generated per unit of compute. And this is precisely where the AI industry is becoming increasingly interesting from an investment perspective, because we are beginning to see the emergence of two powerful forces operating simultaneously: rapid technological improvements are reducing the cost of producing intelligence, while expanding adoption and increasingly sophisticated applications are increasing the quantity of intelligence demanded. The interaction between those two forces could create a much larger market than investors currently anticipate, because every improvement in price-performance can make new categories of work economically viable, while every new category of work can create additional demand for models, data centers, networking, memory, power and specialized hardware. My base case remains that AI consumption will continue growing rapidly as inference becomes cheaper, models become more capable and agents begin performing larger portions of economically useful work, while the physical infrastructure required to support that consumption remains one of the most important bottlenecks in the global technology supply chain. That is why I remain constructive on NVIDIA, TSMC, ASML, Broadcom and the broader AI infrastructure ecosystem, while also recognizing that the ultimate profitability of individual model providers will depend heavily on pricing power, cost efficiency, customer retention, capital allocation and the ability to convert rapidly increasing usage into durable gross profit. As for Anthropic, the latest data make me more optimistic about its commercial momentum, but I would still want to see actual revenue conversion before treating US$100 billion in year-end annualized revenue as a central-case assumption, because the OpenRouter data are powerful evidence of accelerating usage but are not, by themselves, a substitute for consolidated financial disclosure. The company could very well reach that milestone, particularly if enterprise adoption, premium-model usage and agentic workloads continue expanding at anything close to the pace suggested by the latest data. The OpenRouter data certainly make the scenario more plausible, although they should be interpreted as evidence of a powerful demand signal rather than as definitive proof that Anthropic’s total revenue is already accelerating at the same rate. But the bigger conclusion is much more important than whether Anthropic exits December at US$80 billion, US$100 billion or some other annualized figure, because the broader economic implication is that falling prices may be expanding the market rather than destroying it. The world appears to be responding to cheaper intelligence by consuming substantially more of it, as developers use models more frequently, companies deploy them across more departments and increasingly autonomous systems generate workloads that would have been impossible to justify when inference was more expensive. And if that behavior persists, then the economic opportunity in artificial intelligence could continue expanding even as the price of each individual token declines, because the total amount of intelligence consumed may grow faster than the price per unit falls. The market is still debating whether AI infrastructure spending has gone too far, and that debate is entirely reasonable because capital intensity, valuation risk and competitive pressure are all real concerns. Meanwhile, developers are consuming more intelligence, companies are finding new applications for it and demand for computing capacity remains strong, which suggests that the underlying adoption curve has not yet reached the point at which lower prices are causing the market to contract. That is not evidence that every AI stock is attractively valued or that every dollar of capex will earn a good return, because the industry can grow rapidly while investors still overpay for individual assets or companies. But it is powerful evidence that the underlying technology is moving rapidly from experimentation toward large-scale economic deployment, with AI increasingly becoming part of the operating infrastructure of businesses rather than remaining a tool used only for isolated demonstrations. Cheaper intelligence does not necessarily mean a smaller AI market. It may be precisely what creates a much larger one. And that is why I continue to believe the biggest mistake investors can make is confusing falling token prices with falling demand for AI, because the more important question is whether lower prices are unlocking enough additional usage to increase total revenue, compute demand and economic value across the ecosystem. The price of intelligence is collapsing, the consumption of intelligence is exploding and the infrastructure required to produce it is becoming more valuable. Long AI.

AMZNAmazonNeutralAmazonTSMTSMCBullishTSMCAMDNeutralGOOGLAlphabetNeutralAlphabetNVDANVIDIANeutralNVIDIA

3w ago on

$TSM just delivered another ridiculous quarter. Revenue: US$46.7B, +41% YoY Net profit: US$23.6B, +69% YoY Gross margin: 66.4% Operating margin: 58.7% ROE: 45.5% This is now TSMC’s 6th consecutive quarter of record profits. But the most interesting number is elsewhere: wafer shipments grew only 9.6% YoY, while revenue grew 50.9% in NT$. That tells you exactly what is happening. AI does not merely need more chips. It needs far more valuable chips. 2nm + 3nm + 5nm + 7nm already account for 79% of wafer revenue, while HPC has exploded to 68% of total revenue. NVIDIA can win. AMD can win. Google TPU can win. Amazon Trainium can win. Custom ASICs can win. TSMC monetizes the war regardless. And despite massive capex, it still generated NT$477B of quarterly free cash flow. At only 17.1x forward P/E and 11.3x EV/EBITDA, I still think the valuation is extremely attractive for arguably the most important semiconductor manufacturer on Earth. The market keeps debating whether AI capex can generate returns. TSMC is already showing us the answer in its income statement. Long $TSM. One of our core portfolio positions.