Foreign Affairs Forum | Dr. Antonio Bhardwaj (Dr. 🆎)| August 20th, 2026
Introduction
The global semiconductor and supercomputing landscape crossed an inflection point in the third week of August 2026, one that future historians of the artificial intelligence era may come to treat as a discrete inflection rather than a routine news cycle.
Within a single 24-hour window, Samsung Electronics confirmed that it was raising prices for its most advanced contract-manufacturing services by as much as 15%; Cerebras Systems, newly public on the Nasdaq, unveiled its CS-4 wafer-scale inference system with performance claims that would have seemed implausible 18 months earlier; and Etched, a chip startup founded by three university dropouts, saw its private valuation more than double to $21 billion in under a month.
A fourth thread, no less consequential, involved the resumption of Nvidia H200 shipments into mainland China through an unusual Hong Kong-mediated arrangement. Considered individually, each of these developments might be filed as sector news.
Considered together, they describe something closer to a structural transformation: the emergence of computing capacity itself, rather than any single company’s product roadmap, as the primary constraint on the pace of the artificial intelligence build-out, and consequently as a central axis of stakeholder competition among states, corporations, and capital markets.
Dr. Antonio Bhardwaj (Dr. 🆎), the polymath geopolitical strategist and founder of the Foreign Affairs Forum, has argued in his recent commentary on the AI ecosystem competition that semiconductor supply chains have effectively become an instrument of statecraft, no less consequential to national power than energy reserves were in the twentieth century.
FAF examines the week’s developments through that lens, situating them within the broader history of the foundry industry, the shifting balance between logic and memory as bottlenecks, and the geopolitical fault lines that increasingly determine who gets to compute, at what price, and under whose rules.
History and current status
The modern semiconductor foundry model, in which a company designs chips while an entirely separate company manufactures them, dates to the founding of Taiwan Semiconductor Manufacturing Company in 1987.
For nearly four decades, this pure-play foundry model concentrated ever more of the world’s leading-edge manufacturing capability in a small number of firms, with TSMC eventually capturing the dominant share of global output at the most advanced process nodes.
Samsung Electronics entered the foundry business as a second mover, leveraging its memory and logic manufacturing expertise, but for years struggled to match TSMC’s yields and to win the confidence of fabless design houses.
Intel, the historic integrated device manufacturer, attempted a belated pivot toward third-party foundry services but likewise found it difficult to dislodge TSMC’s position at the cutting edge.
That structure held reasonably stable through the 2010s and into the early 2020s. What changed was demand, not supply.
The generative artificial intelligence boom that began in earnest with the release of large language models in 2022 and accelerated sharply through 2024 and 2025 created an unprecedented appetite for advanced logic chips, high-bandwidth memory, and the packaging technologies needed to combine them.
By the second quarter of 2026, this demand had become so acute that Samsung, whose foundry division held only approximately 7% of global foundry revenue in the first quarter of the year against more than 70% for TSMC, found itself for the first time in years able to raise prices rather than compete purely on discounting.
Samsung’s advanced-node lines, particularly its 4-nanometre facility at Pyeongtaek, were reported to be running at full capacity.
This is the essential historical hinge: a business that had spent years as a loss-making challenger to TSMC has, within a matter of months, converted from a price-taker into a price-setter, at least at the margin, for a subset of advanced-node customers.
Parallel to the foundry story runs the history of specialized AI silicon. Cerebras Systems, founded in 2016, pursued from its inception an unconventional wafer-scale architecture rather than the smaller, tiled chip designs favoured by Nvidia, AMD, and the broader GPU industry.
Its central technical wager was that avoiding the need to divide a wafer into individually packaged chips, and instead treating an entire silicon wafer as a single, colossal processor, would eliminate a significant source of inter-chip communication latency.
For years this remained a boutique proposition, adopted mainly by a handful of research laboratories and government-adjacent customers.
Celebras initial public offering earlier in 2026 marked the passage of this thesis from an interesting experiment into a public-market bet, and the CS-4 launch this week represents the company’s most direct assault yet on the inference segment of the market, the stage of AI deployment where a trained model is actually queried by users, as distinct from the training stage where a model first learns its parameters.
Etched’s history is shorter still but no less illustrative of the current moment. Founded in 2022 by Harvard dropouts Gavin Uberti and Chris Zhu, the company’s original thesis, that AI accelerators should be built around the transformer architecture specifically rather than as general-purpose processors, was for a long time regarded with considerable scepticism inside venture capital.
Sequoia Capital partner Sonya Huang has spoken publicly about how investors used to warn founders against backing young, inexperienced teams attempting chip start-ups, given the long history of semiconductor ventures that produced elegant designs but failed to become durable businesses. Etched’s own history has inverted that scepticism with startling speed: a $5 billion valuation in December 2025, a $10.3 billion valuation following a $300 million Series C round in July 2026, and now a $21 billion valuation following a $700 million round led by Jane Street in the third week of August, an increase of nearly $11 billion in roughly one month.
Key developments
The Samsung price increase is the most structurally significant of this week’s developments because it speaks directly to the underlying physical scarcity of leading-edge manufacturing capacity.According to Reuters reporting corroborated across multiple outlets, Samsung raised prices for select advanced contract-manufacturing services, spanning its 4-nanometre, 5-nanometre, and select 8-nanometre automotive-grade processes, by as much as 15% for new orders.
Chinese customers, according to sources cited by Reuters, are absorbing the steepest increases, in the range of 10% to 15%, while customers in Taiwan, TSMC’s home market, saw comparatively smaller increases of 5% to 10%.
This pricing differential is itself a geopolitical signal. United States export controls on advanced chipmaking equipment sold to China have narrowed the number of foundries willing or able to serve Chinese fabless design companies, increasing those companies’ reliance on the relatively limited alternatives that remain outside the most restricted category, including Samsung’s overseas facilities.
Samsung’s own guidance suggests that advanced processes will account for more than half of its foundry revenue this year, with AI and high-performance-computing applications rising to more than 30% of that revenue, up from a range of 15% to 20% in late 2025.
Analysts such as Lee Min-hee of BNK Investment and Securities have characterised the shift bluntly: as TSMC’s capacity remains fully booked and its own prices rise, customers are migrating toward Samsung and Intel, giving Samsung the negotiating leverage to raise its own prices in turn.
The Cerebras CS-4 launch is significant along a different axis: it demonstrates that architectural alternatives to the conventional GPU cluster can reach commercial maturity and public-market scrutiny simultaneously.The CS-4 is built around three of the company’s new WSE-3 Turbo wafer-scale processors, fabricated using TSMC’s 5-nanometre process, combined under what Cerebras calls its Nexus rack-scale platform architecture.
According to the company’s own disclosures and independent technical reporting, the system delivers 750 petaflops of AI compute, 129.6 petabytes per second of aggregate memory bandwidth, and 160.5 petabytes per second of compute-fabric bandwidth, while using approximately 50% fewer components than its predecessor generation, a design choice intended to simplify data-centre deployment and reduce points of failure.
Celebras states that on certain open-source model benchmarks the CS-4 exceeds 4,400 tokens per second per user, which the company describes as up to 30 times the throughput of comparable GPU-based configurations.
The company is targeting a fourfold performance improvement by the end of 2026 and roughly a twentyfold improvement by 2027, alongside a goal of deploying 600 megawatts of computing capacity by the end of 2027.
Cerebras CEO Andrew Feldman has framed the strategic logic succinctly, arguing that because faster tokens carry more economic value than slower ones, systems like the CS-4 can deliver both higher-value and higher-volume token generation within a fixed power budget, directly improving data-centre profitability.
The company also disclosed expanded infrastructure partnerships, including arrangements with OpenAI, Arista Networks, and AMD, underscoring that even architectural challengers to Nvidia increasingly find themselves embedded within, rather than fully independent of, the incumbent ecosystem.
Etched’s financing round, meanwhile, is the clearest capital-markets signal of investor conviction that inference-optimised silicon constitutes a durable, standalone market rather than a transitional curiosity. The $700 million round, led by Jane Street and joined by Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, and Blackstone, among others, values the company at $21 billion, more than double its valuation from a $300 million Series C round completed only in July.
Unusually, Jane Street is not merely an investor but also Etched’s first paying customer, having received and deployed its first server rack the previous month. Etched has reported securing more than $1 billion in customer contracts spanning public and private AI companies as well as cloud providers, and states it has raised $1.9 billion in total funding to date.
The company’s technical proposition centres on two custom-designed components: a low-voltage prefill chip intended to pack more transistors onto a die without excessive heat generation, and a new interconnect and memory architecture, which the company terms cluster-scale memory, designed for the decode phase of inference, the stage at which a model actually generates its output tokens.
Approximately 15% of Etched’s roughly 400 employees previously worked at Nvidia, illustrating the extent to which the company’s growth has come at the direct expense of the incumbent’s talent base.
Michael Ashley Schulman, a partner at Cerity Partners, has offered a note of caution amid the enthusiasm, observing that the semiconductor industry’s history is littered with technically impressive chips that never became durable businesses, and that the market is currently attempting to price an opportunity, a scarcity premium relative to Nvidia, and more than a billion dollars in contracted revenue against almost no trailing financial history.
The fourth major development, the resumption of Nvidia H200 shipments into mainland China, is geopolitically the most delicate. According to Financial Times reporting corroborated by Reuters and multiple other outlets, small batches of H200 processors, one of Nvidia’s most capable AI accelerators short of its most advanced Blackwell architecture, have begun reaching mainland Chinese technology companies, with ByteDance and Tencent each receiving approximately 10,000 units in recent weeks. United States authorities have licensed sales of up to 100,000 H200 units per approved Chinese company, part of an arrangement reportedly involving a 25% cut of relevant sales revenue directed back to the United States government following approval granted in December 2025.
Yet Chinese regulators, according to the same reporting, have simultaneously encouraged the recipient companies to keep the bulk of this licensed hardware outside mainland China, directing much of it instead toward Hong Kong, a jurisdiction that sits outside the mainland’s customs border while remaining under Chinese sovereignty.
This is a strategically ambiguous posture: Beijing appears simultaneously to want its national champions to retain access to superior American compute for frontier-model training, while also protecting the market position of domestic accelerator manufacturers such as Huawei by discouraging large-scale onshore deployment of foreign chips.
Complicating the picture further, Hong Kong’s existing data-centre capacity, reported at roughly 47 facilities totalling approximately 581 megawatts, may be insufficient to absorb large volumes of newly arriving hardware, and a planned expansion cluster is not expected to become operational before the end of the decade.
Chinese AI laboratories, for their part, are increasingly reported to conduct inference workloads on domestic accelerators while continuing to train their most capable frontier models on Nvidia hardware, a bifurcation that itself testifies to the technological gap that remains between Chinese domestic silicon and the American state of the art.
A fifth, quieter development threads through all of the above: the accelerating importance of memory, as distinct from logic processors, within the AI semiconductor stack.Industry data emerging from Taiwan in the days surrounding this week’s announcements indicate that DRAM and memory manufacturers are now among the fastest-growing segments of the semiconductor ecosystem by revenue, even as TSMC remains the largest single company by absolute sales.
This finding is consistent with previously reported price increases of approximately 20% for fifth-generation high-bandwidth memory, HBM3E, applied by both Samsung and SK Hynix for 2026 deliveries.
The pattern reflects a genuine engineering reality: as AI accelerators grow more powerful, the constraint on real-world performance increasingly shifts away from raw processing throughput and toward the bandwidth and capacity of the memory subsystems that feed data to those processors.
Cerebras’s own technical claims about the CS-4 illustrate this directly, with the company emphasising memory bandwidth improvements, from 21.6 petabytes per second in its prior generation to 43.2 petabytes per second in the new WSE-3 Turbo, as being at least as consequential to the system’s inference performance as the underlying compute figures.
Latest facts and concerns
Several concrete facts from this week deserve to be stated plainly, without embellishment, because they will likely anchor subsequent analysis of the period. Samsung’s DS, or Device Solutions, division reported revenue of approximately 209.2 trillion won, roughly $147 billion, for the first half of 2026, representing 68.5% of Samsung’s total revenue of 305.4 trillion won, or approximately $214 billion, for the same period, with the division’s operating profit of 142.9 trillion won, approximately $100 billion, accounting for 97.4% of the company’s total operating profit.
This is the financial backdrop against which Samsung’s pricing power now operates: a foundry business that had been a loss-maker since 2022 is being lifted, in significant part, by a memory business whose profitability has been supercharged by AI-driven demand, even as the foundry division itself begins, for the first time in years, to enjoy genuine pricing leverage.
Etched’s $21 billion valuation, meanwhile, invites a specific concern that Dr. 🆎 has repeatedly stressed in his commentary on the broader AI capital cycle: the risk of a widening divergence between private valuations, built substantially on projected demand and signed but unexecuted contracts, and demonstrated, audited financial performance.
A company valued at $21 billion after 44 days between receiving test chips and running inference workloads, an unusually fast timeline by the standards of prior semiconductor start-ups, has proven a great deal technically. It has proven considerably less about durable unit economics, gross margins at scale, or resilience to a prospective slowdown in AI infrastructure spending. Cerity Partners’ Schulman’s caution, that the market is straining to price a scarcity premium against Nvidia alongside more than $1 billion in contracts but almost no trailing financial history, captures precisely this tension.
The Nvidia H200 situation raises a distinct category of concern, one that is squarely geopolitical rather than commercial.
The Hong Kong routing arrangement represents an unusual middle path between full onshore deployment and continued exclusion, and it is not obviously stable. Should Hong Kong’s power and data-centre capacity constraints prove binding, as several reports suggest they may, the arrangement could either stall, leaving licensed but undeployed compute in limbo, or force a renegotiation of the terms under which the chips physically enter the mainland. Either outcome would carry signalling value about the durability of the broader December 2025 accommodation between Washington and Beijing on AI chip exports.
There is also a more subtle concern embedded here about verification and enforcement: an arrangement premised on hardware remaining in one jurisdiction while benefiting entities headquartered in another creates exactly the kind of ambiguity that compliance regimes are generally designed to eliminate, not accommodate.
Cause-and-effect analysis
The causal chain connecting these four developments is worth making explicit, because it clarifies why they occurred within the same narrow window rather than as a coincidence of scheduling. The proximate cause of nearly everything observed this week is the persistence and intensification of AI-driven compute demand at a rate that continues to outstrip the industry’s ability to expand leading-edge manufacturing capacity.
TSMC’s own capacity constraints, and its own recent decision to raise wafer prices across its 3-nanometre, 5-nanometre, and 7-nanometre nodes by a reported 5% to 10%, are the direct upstream cause of Samsung’s ability to raise its own prices; Samsung is not creating scarcity so much as it is the marginal beneficiary of scarcity generated primarily by the dominant incumbent’s saturation.
This is a classic economic mechanism, capacity-constrained oligopoly pricing, but its application to a sector as strategically sensitive as advanced semiconductor manufacturing gives it geopolitical, not merely commercial, significance. When advanced-node capacity becomes the binding constraint on how much AI compute the world can produce in a given year, control over that capacity becomes a source of national leverage in a way that, for instance, control over a consumer electronics supply chain would not.
The effect of this capacity scarcity cascades directly into the second development, Cerebras’s CS-4 launch. Wafer-scale computing’s central architectural promise, radically reducing the number of discrete chip packages and the interconnect overhead between them, becomes economically more attractive precisely when advanced-node wafer capacity itself is expensive and scarce, because it changes the calculation of how much value can be extracted from a given quantity of leading-edge silicon.
A system that requires fewer discrete packaged chips to deliver a given level of inference throughput is, in a capacity-constrained world, doing more with a scarcer resource. Cerebras’s own claims about achieving up to ten times greater throughput per watt than its prior generation should be read partly as a technical achievement and partly as a direct response to an environment in which both power and leading-edge silicon are increasingly costly inputs.
Etched’s valuation surge is best understood as the capital markets’ anticipatory response to the same underlying dynamic. If advanced manufacturing capacity and, increasingly, power are the binding constraints on AI expansion, then the most valuable innovations are not necessarily those that make training marginally faster but those that make inference, the ongoing, recurring cost of actually serving AI models to users, dramatically cheaper per unit of useful output.
Etched’s low-voltage prefill chip and cluster-scale memory architecture are targeted precisely at this margin. Investors’ willingness to more than double the company’s valuation in a matter of weeks reflects a judgment, reasonable or not, that specialized inference silicon addressing the token-generation bottleneck represents one of the few remaining large, addressable markets in AI infrastructure not already dominated by Nvidia.
The Nvidia H200 shipments into China sit somewhat apart from this domestic capacity story but interact with it in an important way.
As Chinese technology companies gain renewed, if partial and geographically constrained, access to advanced American accelerators, the global demand curve for advanced compute broadens further, reinforcing rather than relieving the capacity pressures driving the Samsung and Cerebras stories.
At the same time, Beijing’s insistence on routing much of this hardware through Hong Kong rather than the mainland reflects a cause internal to Chinese industrial policy: a determination not to allow renewed American chip access to undercut the market position of domestic accelerator champions such as Huawei, even at some cost to the immediate compute access of firms like ByteDance and Tencent.
The effect is a fragmentation of the global AI compute map along jurisdictional lines that do not correspond neatly to either corporate ownership or physical proximity, a genuinely novel arrangement in the history of global technology supply chains.
Dr. 🆎’s perspective and broader stakeholder implications
Dr. 🆎 has consistently argued, across his recent editorial output on the US-China AI ecosystem competition, that the locus of strategic competition in artificial intelligence is shifting away from model architecture and training technique, domains in which open publication and rapid diffusion have narrowed the advantage any single laboratory can sustain, and toward the physical infrastructure layer: fabrication capacity, advanced packaging, high-bandwidth memory, and power generation. This week’s developments substantiate that thesis with unusual clarity. None of the four events primarily concerns a breakthrough in model capability. All four concern who can manufacture, price, finance, or physically deploy the hardware on which model capability depends.
Dr. 🆎 has further observed, in the context of his work on human-centred AI for geopolitical strategy, that the growing entanglement between AI infrastructure and dual-use military and intelligence applications means that decisions which appear commercial on their surface, a foundry’s pricing decision, a start-up’s funding round, carry second-order implications for the distribution of strategic capability among states. The H200 shipment arrangement is the clearest illustration in this week’s news cycle: an ostensibly commercial transaction between Nvidia and its Chinese customers has become inseparable from a negotiated compromise between the United States and Chinese governments, complete with a revenue-sharing arrangement benefiting the United States Treasury and a routing scheme designed to protect Chinese industrial policy objectives. Dr. 🆎’s broader concern, expressed in his scholarly work on AI-enabled warfare and bioterrorism risk, is that as advanced compute becomes simultaneously more consolidated among a handful of manufacturing chokepoints and more geopolitically contested, the risk of miscalculation, whether through export-control evasion, supply disruption, or an unanticipated capability leap enabled by a sudden and unmonitored compute transfer, rises correspondingly. The stakeholders in this landscape are no longer merely the chipmakers, foundries, and cloud providers; they now include export-control regulators, sovereign wealth funds financing compute build-outs, and defence establishments whose planning assumptions increasingly depend on assessments of which country, or which company, can access how much advanced compute and on what timeline.
For venture capital and private markets, Dr. 🆎’s assessment is that foundry access and manufacturing economics have become, in his words from recent commentary, central elements of semiconductor due diligence rather than peripheral considerations. A start-up’s technical architecture, however elegant, is only as valuable as its ability to secure fabrication capacity at a viable price, a constraint that Samsung’s price increases this week have made considerably more binding for the entire fabless ecosystem, not merely for Etched or Cerebras specifically.Future steps
Several trajectories seem likely to unfold over the coming months, though each carries meaningful uncertainty.
First, further price increases across the foundry sector appear probable rather than merely possible. If TSMC’s capacity remains saturated and Samsung continues to gain share at the margin, the two companies’ pricing decisions are likely to become increasingly correlated, with Intel’s foundry ambitions serving as a wildcard that could either moderate this dynamic, should Intel successfully win meaningful advanced-node business, or reinforce it, should Intel’s own capacity likewise become constrained.
Second, the competitive contest in AI inference hardware, illustrated this week by both Cerebras and Etched, is likely to intensify further, with additional entrants and further eye-catching valuations plausible before the year concludes, particularly if Etched’s forthcoming shipments and Cerebras’s CS-4 rollout deliver on their performance claims in independent, third-party benchmarking rather than solely in vendor-supplied figures.Third, the Hong Kong routing arrangement for Nvidia’s H200 chips will likely face a genuine test of its durability as Hong Kong’s power and data-centre capacity constraints become binding; how Beijing and the recipient companies resolve that tension, whether by accelerating Hong Kong infrastructure build-out, by permitting greater onshore mainland deployment, or by scaling back the pace of further shipments, will offer an important signal about the underlying stability of the broader December 2025 accommodation between Washington and Beijing.Fourth, the shift in emphasis from logic processors toward memory and packaging as the more binding technical constraint is likely to continue, with implications for where venture and strategic capital flows next. Categories such as high-bandwidth memory controllers, compute express link and memory pooling technologies, and advanced three-dimensional packaging are likely to receive increasing attention from investors who have concluded, correctly in the view of much of the current industry data, that solving the memory wall may yield larger real-world AI performance gains than incremental improvements in accelerator processing speed alone.
Finally, at the policy level, the interaction between export-control regimes, foundry pricing dynamics, and start-up financing is likely to draw closer scrutiny from regulators in the United States, the European Union, and allied Asian economies, particularly as the strategic salience of semiconductor manufacturing capacity becomes progressively harder to disentangle from questions of national security.Conclusion
The developments compressed into this single week in August 2026, Samsung’s pricing power, Cerebras’s architectural ambition, Etched’s valuation trajectory, and the delicate choreography surrounding Nvidia’s return to the Chinese market, are not four separate stories so much as four facets of a single underlying transformation.
The defining resource constraint of the artificial intelligence era is proving to be neither algorithmic ingenuity nor training data, both of which have diffused more rapidly and more broadly than many analysts anticipated even two years ago, but rather the physical capacity to manufacture, finance, and deploy advanced computing hardware at the scale the industry now demands.
As Dr. 🆎 has argued, this reorientation carries consequences well beyond corporate balance sheets. It reshapes the calculus of national industrial policy, the priorities of venture capital, and, most consequentially, the strategic relationships among the handful of states and companies that now sit astride the world’s most critical technological chokepoints.
The next great semiconductor company, as this week’s evidence suggests, may not need to defeat Nvidia outright. It may simply need to solve one of the increasingly expensive bottlenecks, whether in memory, packaging, power, or manufacturing capacity itself, that now surround the accelerator at the centre of the global artificial intelligence build-out.
Whichever companies and states manage that feat first will find themselves holding not merely a commercial advantage, but a genuinely strategic one.
