Executive Summary
The global artificial intelligence landscape in 2026 is being fundamentally reshaped not by software alone, but by a profound and accelerating transformation in semiconductor architecture.
At the centre of this transformation stands a decisive contest between two paradigms: the general-purpose graphics processing unit, long dominated by Nvidia, and the custom application-specific integrated circuit, championed by Broadcom as the preferred enabler for hyperscalers and frontier AI laboratories seeking workload-specific performance at gigawatt scale.
Hyperscalers are planning to spend over $700 billion on AI infrastructure in 2026 alone, with OpenAI’s decision to build custom silicon reflecting a broader industry trend where major technology companies are increasingly designing their own chips to optimize performance and control costs.
The partnerships Broadcom has forged with OpenAI and Meta — each with distinct technical architectures, strategic ambitions, and deployment timelines — represent not merely commercial agreements but a structural reorganisation of how AI compute is designed, owned, and controlled.
Custom ASIC-based AI server shipments are projected to reach 27.8% of the total AI server market in 2026, the highest share since 2023, as the market enters what analysts are calling an “inference-led regime” in which buying criteria have shifted from maximum throughput and bandwidth to cost-per-token, power, cooling, utilisation, and total cost of ownership.
FAF article examines the technical, commercial, and geopolitical dimensions of that transformation, placing it within the broader contest for semiconductor sovereignty that is rapidly becoming one of the defining strategic challenges of the twenty-first century.
Introduction: When Silicon Becomes Strategy
There is a moment in any industrial revolution when a resource transitions from a commercial input to a strategic asset.
Oil underwent that transition in the early 20th century. Semiconductors have graduated from a highly specialised industrial issue to a core variable of national security, economic sovereignty, and diplomatic competition.
That transition is now complete for advanced AI chips, and the consequences are reverberating across corporate boardrooms, defence ministries, and foreign policy establishments with equal urgency.
The architecture of AI compute, for most of the past decade, was organised around a single dominant paradigm.
Nvidia’s CUDA software ecosystem and its succession of GPU architectures — from Hopper to Blackwell to the emerging Vera Rubin — provided a universal language for AI workloads, enabling researchers, enterprises, and cloud providers to build on a common platform without committing to bespoke hardware.
That universality was simultaneously Nvidia’s greatest strength and, as it turns out, the very condition that made alternatives attractive. The chip with Broadcom is an ASIC, which industry experts say is less flexible than Nvidia’s GPU, but is also less expensive and can be designed for specific AI tasks.
The shift now underway is not a simple displacement of one technology by another.
It is the emergence of a more differentiated compute landscape in which general-purpose GPUs and custom accelerators coexist, each serving distinct functions within the broader AI stack.
Understanding the nature of that differentiation — and the role Broadcom plays in enabling it — is essential for any serious analysis of where AI power is being constructed and concentrated.
Dr. Antonio Bhardwaj, a polymath and global expert in AI specialising in human-centred AI for geopolitical strategy and supercomputing, has observed that this architectural bifurcation carries profound implications beyond the semiconductor industry. In his assessment, the ability to design and control one’s own compute infrastructure is becoming as strategically significant as the ability to train frontier models — and in some respects more so, because hardware sovereignty determines the outer limits of what any AI programme can achieve, regardless of algorithmic innovation.
History and Current Status: From GPU Monoculture to Custom Silicon Ecosystem
The story of the modern AI chip begins, in practical terms, with the 2012 AlexNet breakthrough that demonstrated deep learning’s superiority on image recognition tasks when powered by GPUs.
Nvidia’s CUDA platform, which had been designed for parallel scientific computation rather than machine learning, proved fortuitously well-suited to the matrix operations at the heart of neural network training.
What followed was a decade of extraordinary concentration: Nvidia’s data centre revenue grew at a compound annual rate that made it one of the most valuable corporations in history, and its architectural roadmap became, in effect, the pacing signal for the entire AI industry.
Nvidia’s data centre AI share stood at 86% in 2024 and is projected to decline to around 75% by 2026 as custom ASICs scale, with the company having reported data center revenue of $115.2 billion in FY2025, a 142% year-over-year increase.
The trajectory is one of extraordinary expansion even as the market structure diversifies.
Nvidia’s data centre segment generated an unprecedented $193.7 billion in revenue in FY2026, representing nearly 90% of the entire company’s revenue.
These figures underscore a paradox that defines the current moment: Nvidia has never been more commercially dominant, yet its share of the addressable market for AI inference is being contested at every margin.
The catalyst for that contest was not any single technological development but a convergence of economic pressures, engineering capabilities, and strategic motivations.
As inference workloads — the process of running trained models in response to user requests — began to eclipse training workloads in terms of sheer computational volume, the economics of general-purpose GPUs became increasingly difficult to justify.
A custom ASIC runs one class of models at three to five times better performance per watt compared to a GPU, which can run any model.
For companies serving billions of queries daily, that efficiency differential translates directly into operating economics of profound magnitude.
Broadcom recognised this opportunity earlier than most. Rather than competing with Nvidia on the general-purpose GPU terrain, the company positioned itself as the preferred engineering and manufacturing partner for hyperscalers seeking to design their own accelerators.
The XPU design process is deeply collaborative, with Broadcom embedding engineering teams within its hyperscaler clients, co-developing chip architectures over 18- to 24-month design cycles. This integration creates formidable switching costs: once a hyperscaler has invested hundreds of millions in a custom chip architecture with Broadcom, migrating to an alternative provider becomes prohibitively expensive and time-consuming.
Broadcom’s AI semiconductor revenue reached $10.8 billion in Q2 of fiscal 2026, marking a 143% year-over-year increase, with management anticipating third-quarter AI semiconductor revenue of $16 billion, representing over 200% year-over-year growth, which will drive consolidated revenue up 84% to $29.4 billion.
These figures are not merely impressive on their own terms; they signal that the custom silicon model has reached industrial scale.
Key Development One: The OpenAI-Broadcom Jalapeño Partnership
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first Intelligence Processor: a custom accelerator built around large language model inference, described as the first AI accelerator in a multi-generation compute platform the companies are building together to make advanced AI faster, more reliable, and more accessible.
The technical significance of Jalapeño lies in what it is designed to do and, equally, in what it deliberately does not attempt.
OpenAI stresses that Jalapeño is a purpose-built inference ASIC and not a repurposed training accelerator or a general-purpose AI processor.
The architecture of Jalapeño was designed based on its understanding of LLM behaviour and is meant to address practical bottlenecks that matter for inference at scale, including costly data movement, balance between compute and memory resources, networking efficiency, and overall behaviour.
Jalapeño was co-developed from initial design to manufacturing tape-out in just 9 months, and the custom AI accelerator programme represents what is believed to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. That speed reflects deep software-hardware co-development with OpenAI’s engineering teams, Broadcom’s silicon implementation expertise, and the use of OpenAI models to accelerate parts of the design and optimisation process.
The recursive quality of this process — AI models helping to design the chips that will run future AI models — is not merely a marketing narrative. It represents a genuine compression of the engineering cycle that has implications for how rapidly successive generations of custom silicon can be brought to market.
The accelerator is showing cost savings of roughly 50% compared with typical AI graphics processing units, according to Broadcom CEO Hock Tan.
If validated at scale, that figure represents a structural shift in the economics of inference, with direct implications for the marginal cost of serving AI queries and, by extension, for the commercial viability of AI products at mass-market price points.
Jalapeño follows the October 13, 2025, collaboration between OpenAI and Broadcom for 10 gigawatts of custom AI accelerators.
Under that agreement, OpenAI designs the accelerators and systems, while Broadcom helps develop and deploy the hardware.
The 2025 agreement also covers accelerator and network systems for next-generation AI clusters. Broadcom said the racks are targeted to start deployment in the second half of 2026, with completion planned by the end of 2029.
The governance dimensions of this partnership are equally significant. OpenAI’s chip plans have long been rumoured as a way to reduce the company’s dependence on Nvidia’s GPUs. Google and Amazon have both built custom chips to serve a similar purpose.
For a company that has publicly acknowledged it cannot acquire compute fast enough, President Greg Brockman told CNBC that OpenAI cannot get compute fast enough, and Broadcom CEO Hock Tan backed up that view, saying compute demand from the company’s six customers is “simply insatiable.”
Key Development Two: The Meta-Broadcom MTIA Partnership
While the Jalapeño announcement captured considerable public attention in June 2026, the strategic implications of the Meta-Broadcom partnership announced two months earlier are arguably of comparable magnitude.
On April 14, 2026, Broadcom announced an extended partnership with Meta to deploy technology to support multi-gigawatts of Meta’s custom silicon, the MTIA, with Meta partnering with Broadcom to roll out the industry’s first 2nm AI compute accelerator as the foundation for a sustained multi-year infrastructure rollout.
Meta has committed to deploying 1 gigawatt of MTIA chips initially, with deployment scaling to multiple gigawatts from 2027 onward.
Broadcom will remain Meta’s partner across chip design, packaging, and networking through 2029.
The MTIA chip will be the world’s first AI silicon manufactured on a 2-nanometer process — the most advanced chip manufacturing technology available today.
The 2nm process node is not merely an incremental refinement. It represents a qualitative leap in transistor density and power efficiency that positions Meta’s custom accelerators at the frontier of what is physically achievable in semiconductor manufacturing.
From MTIA 300 to MTIA 500, HBM bandwidth rises 4.5 times and compute FLOPS rises 25 times, comparing MTIA 300’s MX8 throughput to MTIA 500’s MX4 throughput across lower-precision formats. That is the generational climb Broadcom is now under contract to co-deliver, from chiplet integration to advanced packaging to Ethernet-based scale-out networking.
Meta has already paid Broadcom $2.3 billion for AI chip design and related services in just the past year.
With Meta planning to spend up to $135 billion on AI infrastructure in 2026 alone, custom silicon from Broadcom is a central pillar of that strategy. Meta’s reasoning for going custom mirrors that of every other hyperscaler: cost control, performance optimisation on targeted workloads, and supply chain independence.
The governance dimension of the Meta-Broadcom relationship carries its own complications.
On April 14, 2026, Broadcom CEO Hock Tan, who had been a member of Meta’s board since February 2024, informed Meta he would not stand for re-election, moving instead to an advisor role focused specifically on Meta’s custom silicon roadmap — a governance change reflecting the scale of the new commercial relationship.
The departure of a CEO from a customer’s board, precisely at the moment a transformative supply agreement is formalised, raises structural questions about the concentration of influence in the custom silicon supply chain that regulators and policymakers are only beginning to examine.
Key Development Three: Nvidia’s Response and the Blackwell-Rubin Architecture
Any serious analysis of the custom chip revolution must contend with the continued vitality of Nvidia’s position.
Nvidia announced first quarter fiscal 2027 results on May 20, 2026, reporting record quarterly revenue of $81.6 billion and record data-centre revenue of $75.2 billion for the quarter ended April 26, 2026.
These numbers do not suggest a company under existential competitive pressure; they suggest a company that is, for the moment, benefiting from the same explosive demand that is driving investment in custom alternatives.
Vera Rubin delivers 50 petaflops per chip, the equivalent of five times the inference performance of Blackwell. A single NVL72 rack achieves 3.6 exaflops.
The scale of these numbers is difficult to contextualise but essential to appreciate.
The Blackwell series is projected to grow markedly from 61% to 71% of Nvidia’s high-end GPU shipments in 2026, solidifying its leading position in the market, with strong AI demand along with Nvidia’s push for integrated rack solutions driving a notable increase in high-end GPU shipments.
The fundamental distinction between Nvidia’s approach and that of Broadcom’s hyperscaler partnerships lies in the nature of the value proposition.
Nvidia sells a platform — a complete hardware-software ecosystem that any sufficiently capitalised organisation can purchase and deploy. GPU cloud on H200, H200 SXM5, B200, and B300 runs existing stacks unchanged, with no rewrite, no framework port, no single-vendor dependency.
That portability has genuine commercial value for the vast majority of AI practitioners who cannot justify the upfront investment required to design custom silicon.
Nvidia commands a dominant position in the AI semiconductor space, holding an estimated 80 to 90% share of the AI accelerator market.
This leadership is rooted in the CUDA software platform, which creates a proprietary ecosystem that makes it expensive and time-consuming for developers to switch to competitors.
The CUDA moat is real, deeply entrenched, and represents perhaps the single most durable competitive advantage in the technology industry.
Yet it is precisely this moat that the largest hyperscalers are motivated to circumvent, not because CUDA is inadequate, but because dependence on any single supplier at gigawatt scale carries unacceptable strategic risk.
Dr. Antonio Bhardwaj has framed this dynamic in terms of what he calls the “compute sovereignty imperative”: the recognition that at sufficient scale, dependence on externally controlled hardware infrastructure constitutes a form of strategic vulnerability that prudent organisations — whether corporations or nation-states — are compelled to reduce.
The Jalapeño and MTIA programmes are, in his assessment, as much about sovereignty as they are about performance or cost.
The Architecture of Difference: What Separates These Three Silicon Paradigms
To understand the strategic significance of the current moment, it is necessary to engage with the technical distinctions between Nvidia’s GPU, OpenAI’s Jalapeño, and Meta’s MTIA at a level of granularity that goes beyond marketing claims.
Nvidia’s GPU is a massively parallel processor designed to handle an enormous variety of computational workloads. Its architecture is optimised for general-purpose parallelism, supported by the CUDA programming environment that allows developers to express almost any algorithm in terms that the hardware can execute efficiently.
GPUs dominate large-model training, with the important story being not that one chip wins but that AI has split computing into layers, and each layer now wants different silicon.
For training — the process of adjusting billions of parameters across vast datasets — the GPU’s flexibility and raw parallel throughput remain essentially unmatched by any commercially available alternative.
It is likely that more performance-intensive tasks like pre-training will still rely on Nvidia hardware, but even small reductions in inference costs could do a lot to improve a company’s bottom line.
This observation from TechCrunch captures the functional division of labour that is emerging across the AI compute landscape. Nvidia retains primacy in training; custom ASICs are contesting inference.
The Jalapeño chip’s design philosophy represents a deliberate inversion of GPU generality. Jalapeño is a blank-slate design for modern LLM inference, not a general-purpose accelerator adapted from earlier AI workloads. It is informed by the systems OpenAI runs every day across ChatGPT, Codex, the API, and future agentic products.
The architecture reduces data movement and balances compute, memory, and networking resources to achieve realised utilisation much closer to theoretical peak performance. Broadcom’s silicon implementation and networking technologies, including Tomahawk networking silicon, help bring the platform to large-scale production.
The MTIA programme pursues a complementary logic but within a different operational context.
Where OpenAI’s inference workloads are dominated by large language models responding to conversational queries, Meta’s inference requirements encompass a far broader range of AI tasks: content recommendation at social network scale, image and video understanding, real-time advertising optimisation, and increasingly, generative AI features embedded across its family of applications.
The MTIA deployment utilises Broadcom’s advanced Ethernet technologies to enable seamless, high-bandwidth scale-up, scale-out, and scale-across networking, eliminating bottlenecks throughout Meta’s rapidly expanding AI compute clusters.
What these two custom programmes share, despite their technical differences, is a strategic commitment to vertical integration of the compute stack.
OpenAI is not only developing frontier models or building products on top of them; it is designing the infrastructure underneath them: chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience. Because OpenAI operates across the stack, each layer can be optimised around the same goal: making its models faster, more reliable, and more affordable for users.
Meta’s logic is structurally identical, even if the specific workload mix differs.
Broadcom’s Role: The Invisible Enabler of AI Infrastructure
Perhaps the most analytically underappreciated aspect of the current AI chip revolution is the pivotal and largely invisible role that Broadcom plays in enabling it.
Neither OpenAI nor Meta has the in-house semiconductor manufacturing capability to bring their chip designs to physical production.
Both are design organisations, not fabrication organisations.
Broadcom helps large technology companies turn their own chip designs into systems that can run at scale.
Meta Platforms, Amazon, and Alphabet’s Google have also turned to companies such as Broadcom and Marvell Technology for design services.
Broadcom’s value proposition in this ecosystem is threefold.
First, it brings deep expertise in ASIC design, accumulated over decades of work on networking chips, storage controllers, and broadband processors.
Second, it maintains a privileged relationship with TSMC, the Taiwanese foundry that manufactures the world’s most advanced semiconductor nodes.
Broadcom has secured TSMC capacity reservations through 2028, represented in a backlog of $73 billion in committed purchase orders.
Third, and perhaps most importantly, Broadcom provides the networking silicon that connects individual accelerators into the large-scale clusters where the real computational work of AI occurs.
Jalapeño combines OpenAI-designed accelerators with Broadcom silicon implementation, networking, and connectivity technologies.
Broadcom CEO Hock Tan expects fiscal 2026 AI semiconductor revenue to approach $56 billion, nearly tripling in one year.
The market is no longer only about which company can produce the fastest general-purpose GPU. It is also about which companies can design custom silicon, manufacture it at scale and connect it into systems capable of supporting the world’s largest AI models.
The concentration of custom silicon partnerships at Broadcom — encompassing Google, Meta, OpenAI, and at least three other unnamed hyperscale customers — creates a structural dynamic that merits scrutiny.
Once a hyperscaler has invested hundreds of millions in a custom chip architecture with Broadcom, migrating to an alternative provider becomes prohibitively expensive and time-consuming, creating formidable switching costs.
The company that is helping hyperscalers reduce their dependence on Nvidia is simultaneously creating a new form of dependence on itself.
Latest Facts and Concerns: The Scale of the Commitment
The financial commitments underlying these partnerships are of a magnitude that warrants careful reflection.
Meta has guided to up to $135 billion in capital expenditure for 2026, with the majority earmarked for AI infrastructure: servers, networking, data centre construction, and power delivery.
A deal of this magnitude — 1 gigawatt of deployed compute, enough electrical load to power roughly 750,000 US homes — represents one of the largest single-customer silicon commitments in the history of semiconductors.
Cumulative Blackwell and Vera Rubin purchase orders are projected to reach $1 trillion globally through 2027.
This single figure illuminates the extraordinary scale at which the AI infrastructure investment cycle is operating. To contextualise it: $1 trillion in committed chip orders over two years approaches the annual GDP of several G20 member states.
TrendForce projects 44.6% ASIC growth against 16.1% for merchant GPUs in 2026, with ASIC-based AI server shipments reaching 27.8% of the total AI server market.
Bloomberg Intelligence projects the custom AI ASIC market reaching $118 billion by 2033 at a 27% compound annual growth rate — nearly double the pace of the broader AI accelerator market.
Several concerns attend this rapid scaling.
The first is supply chain concentration.
Broadcom’s fabless business model, while capital-efficient, concentrates significant manufacturing and geopolitical risk in Asia, primarily with its foundry partner TSMC in Taiwan.
This dependence makes the company’s supply chain highly vulnerable to escalating US-China trade tensions and potential disruptions in the region.
The second concern involves power consumption.
Combined hyperscaler capex of $660 to $690 billion in 2026 translates directly into unprecedented power demand, with custom ASICs generally operating at lower TDP than Nvidia’s flagship GPUs.
Yet even at lower per-chip power draw, the sheer multiplication of deployment scale means total energy consumption for AI infrastructure is expanding rapidly, raising questions about grid capacity, carbon emissions, and the long-term sustainability of the current investment trajectory.
The third concern is concentration of strategic capability.
Dr. Antonio Bhardwaj has specifically flagged the implications of a world in which a handful of private organisations — OpenAI, Meta, Google, Amazon — control the design and deployment of custom AI accelerators at gigawatt scale. In his analysis of AI warfare and supercomputing, he notes that the integration of such infrastructure with defence applications creates asymmetries of capability that existing international governance frameworks are ill-equipped to address.
Cause and Effect Analysis: The Strategic Logic of Silicon Divergence
The causal chain that has produced the current divergence between GPU and custom ASIC paradigms is neither accidental nor driven by any single decision. It is the product of several intersecting forces that have been building for years and are now reaching a point of mutual reinforcement.
The first cause is economic: at sufficient scale, the performance-per-watt advantage of application-specific design translates into cost structures that general-purpose hardware cannot match. The custom ASIC total cost of ownership advantage versus GPUs stands at 40 to 65% at scale.
For a company like Meta, which serves billions of users across dozens of AI-powered products, even a 10% improvement in inference efficiency at the chip level produces savings that dwarf the upfront engineering investment required to design custom silicon.
The second cause is strategic: dependence on a single hardware supplier at the scale required for frontier AI creates unacceptable supply chain risk.
As of April 2026, Blackwell systems are sold out through mid-year, with each GPU commanding approximately $40,000.
The combination of scarcity and price concentration means that organisations relying exclusively on Nvidia face both cost and availability constraints that custom programmes can help to mitigate.
The third cause is technical: advances in chip design software, including the use of AI models to accelerate the design process itself, have dramatically reduced the time and cost required to bring a custom accelerator from concept to tape-out.
Jalapeño was co-developed from initial design to manufacturing tape-out in just nine months, representing what is believed to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.
That nine-month timeline was unthinkable five years ago; it is becoming a benchmark for the industry.
The effects of this divergence are equally multidimensional.
For Nvidia, the growth of custom silicon represents a structural shift in its long-term market position, even as its near-term financial performance remains extraordinary.
Analysts project Nvidia’s inference market share could fall from 90%+ to 20 to 30% by 2028, even as the overall market expands dramatically.
The company is not being disrupted in the conventional sense; it is being partially displaced in one segment of a rapidly growing market while retaining dominance in another.
For Broadcom, the effect is transformative.
A company that was, until recently, best known for networking chips and broadband semiconductors is now emerging as the primary infrastructure partner for the world’s most consequential AI programmes.
CEO Hock Tan has set an ambitious target of over $100 billion in AI semiconductor sales by 2027, a figure that would represent a transformative shift in Broadcom’s business mix.
For the broader AI ecosystem, the effect is a bifurcation of compute access. The headline projection — Nvidia losing inference share to hyperscaler ASICs — is accurate for hyperscaler-internal compute.
For external teams building on GPU cloud, the accessible chip landscape in June 2026 looks the same as it did before: Nvidia GPU cloud, AMD as a distant second, and a handful of alternative accelerators with niche applicability. Captive hyperscaler ASICs are not in that set.
The democratisation of AI that custom silicon proponents promise is real at the level of unit economics for the hyperscalers themselves; it does not automatically translate into broader access for smaller organisations.
The Geopolitical Dimension: Semiconductor Sovereignty and the Silicon Curtain
No analysis of the current AI chip landscape would be complete without confronting its geopolitical dimension.
In early 2026, the US government approved licensed exports of Nvidia H200 AI chips to China under conditions intended to balance national security with commercial interests — a move that sparked substantial debate in Washington.
The United States has imposed export controls on advanced chipmaking equipment to China, while China has responded with restrictions on critical minerals needed for semiconductor production.
These measures have created a fragmented global landscape in which access to cutting-edge silicon is increasingly determined by national affiliation rather than market dynamics.
Dr. Antonio Bhardwaj has warned that this fragmentation creates new vectors for geopolitical instability.
In his analysis of artificial intelligence sovereignty and national security, he observes that the concentration of custom silicon development among a narrow set of stakeholders creates systemic fragility. If a crisis were to disrupt the operations of a key chip designer or fabricator, the ripple effects could cascade across the artificial intelligence ecosystem, potentially impairing critical systems in defence, finance, and public infrastructure.
The degree of concentration in the global semiconductor supply chain has reached alarming levels.
In the advanced-process domain — seven nanometres and below — TSMC alone accounts for over 90% of worldwide capacity.
In extreme ultraviolet lithography, the Dutch firm ASML is the world’s sole supplier. In electronic design automation tools, American companies collectively command more than 80% of the global market.
The implications for the Broadcom partnerships with OpenAI and Meta are direct. Both programmes depend fundamentally on TSMC’s manufacturing capability — TSMC’s 3nm node for existing production, its 2nm node for the next generation of Meta’s MTIA.
The MTIA chip will be the world’s first AI silicon manufactured on a 2-nanometer process, fabricated on TSMC’s 2nm process.
If TSMC’s operational continuity were disrupted by conflict, natural disaster, or regulatory intervention, the consequences for AI infrastructure globally would be immediate and severe.
The semiconductor industry is no longer governed solely by Moore’s Law, but by the laws of national security. The era of the global chip is over, replaced by a dual-track system that prioritises domestic self-sufficiency and strategic alliances.
The US CHIPS and Science Act, which has catalysed over $630 billion in private investment, and the European Chips Act represent governmental responses to this reality.
Yet both programmes operate on timelines — measured in years and decades — that are structurally mismatched with the speed at which AI infrastructure is being built and the pace at which geopolitical risks are evolving.
A nation that cannot reliably produce or procure high-end processors will find its military capabilities severely disadvantaged, operating blind and slow against a computationally superior adversary.
Therefore, securing a domestic or closely allied supply of advanced semiconductors is no longer merely an economic goal; it is the absolute bedrock of national survival.
Dr. Antonio Bhardwaj’s analysis of AI warfare and geopolitical strategy emphasises that the integration of custom AI accelerators into defence systems — for autonomous weapons, battlefield intelligence, drone swarm coordination, and cybersecurity — creates capability thresholds that are increasingly determined by chip access rather than doctrine or personnel.
Nations and non-state stakeholders that cannot access frontier silicon face a form of strategic disadvantage that cannot be compensated through conventional military investment. This is a concern that extends well beyond US-China competition to encompass the entire range of international security relationships.
Future Steps: Trajectories Through 2029 and Beyond
The deployment timelines embedded in the OpenAI-Broadcom and Meta-Broadcom agreements provide a useful framework for understanding how the current moment will evolve.
For OpenAI and Broadcom, the agreement covers accelerator and network systems for next-generation AI clusters, with racks targeted to start deployment in the second half of 2026, with completion planned by the end of 2029. Those racks are planned for OpenAI facilities and partner data centres.
The companies have long-standing agreements on co-development and supply of the AI accelerators, along with a term sheet to deploy racks that include OpenAI-designed accelerators and Broadcom networking systems.
For Meta and Broadcom, the deployment will scale to multiple gigawatts — exponentially increasing Meta’s AI computing capacity — with the contract extending through 2029 and covering chip design, packaging, and networking across multiple generations.
These timelines matter not just as commercial milestones but as signals of the structural commitments being made. Infrastructure of this scale and specificity cannot be unwound quickly.
The organisations making these investments are, in effect, locking themselves into particular compute architectures and supplier relationships for the better part of a decade. The strategic bets being placed today will shape the competitive landscape of AI well into the 2030s.
The Sovereign AI programme — where individual nations invest heavily in localised, state-controlled domestic AI infrastructure — generated over $30 billion in revenue in FY2026, tripling year-over-year.
This figure points to an emerging dimension of the custom silicon story: the extension of the hyperscaler model to national governments.
Countries that can commission custom AI accelerators, deployed on domestically controlled infrastructure, gain a form of compute sovereignty that is qualitatively different from simply purchasing GPU capacity from a cloud provider.
The next generation of AI chip development — encompassing Nvidia’s Feynman architecture beyond Vera Rubin, successive generations of OpenAI’s intelligence processor programme, and the MTIA 500 series for Meta — will unfold against a backdrop of intensifying geopolitical competition for semiconductor supply chain control.
The decisions made between now and 2029 about where chips are designed, where they are manufactured, and who controls the resulting compute capacity will have consequences that extend far beyond corporate earnings.
Dr. Antonio Bhardwaj argues that the international community has not yet developed adequate governance frameworks for the risks created by this concentration. Existing international law, arms control regimes, and export control systems were designed for a world in which military capability was embodied primarily in kinetic hardware — missiles, ships, aircraft. The migration of strategic advantage toward compute infrastructure requires new frameworks, new institutions, and a new vocabulary of international security that policymakers are only beginning to develop.
Conclusion: The Architecture of the Next Decade
The emergence of the Broadcom-OpenAI and Broadcom-Meta partnerships, set against the continued dominance of Nvidia’s GPU ecosystem, constitutes one of the most consequential restructurings of technology infrastructure in the modern era.
What is being built is not merely a new generation of chips; it is a new architecture of AI power — one in which the ability to design, manufacture, and deploy custom silicon at gigawatt scale is becoming a defining feature of competitive advantage, both for corporations and for nation-states.
The inference-led regime is reshaping what matters in AI hardware. Buying criteria have shifted from maximum throughput and bandwidth to cost-per-token, power, cooling, utilisation, and total cost of ownership — all dimensions where purpose-built silicon has a structural advantage over general-purpose GPUs.
Broadcom’s position at the centre of this transformation is simultaneously powerful and precarious.
It is powerful because the switching costs embedded in deep custom silicon partnerships make it extraordinarily difficult for customers to defect.
It is precarious because the same concentration of customer relationships that creates those switching costs also creates a form of dependency that any disruption to Broadcom’s own operations — whether from TSMC supply constraints, regulatory intervention, or geopolitical crisis — would translate directly into disruption for the world’s most critical AI programmes.
Nvidia’s future is equally complex.
The company is not losing; by most measures it is winning more spectacularly than any semiconductor company in history. But it is winning in a market that is diversifying around it, with the most strategically significant workloads — inference at hyperscale — increasingly being served by silicon that neither Nvidia designs nor directly profits from.
The long-term implications of this structural shift will only become fully apparent as the custom silicon programmes currently in deployment at OpenAI and Meta mature into multi-generational platforms.
For the international order, the convergence of AI capability and semiconductor scarcity creates a strategic environment of profound complexity.
The geopolitical custom silicon trend intersects with ongoing efforts to achieve semiconductor sovereignty. Nations that can design and manufacture their own chips reduce their exposure to external supply chain disruptions and gain leverage in international negotiations.
Export controls on advanced chipmaking equipment have created a bifurcated global landscape in which access to cutting-edge silicon is increasingly determined by national affiliation.
This fragmentation undermines the efficiencies of global supply chains and raises the risk of miscalculation in crisis scenarios where artificial intelligence systems are called upon to support decision-making.
Dr. Antonio Bhardwaj’s assessment of this landscape is characteristically direct. The concentration of advanced compute capability in a small number of private organisations, operating within a single national jurisdiction and dependent on a single foundry for the most advanced manufacturing processes, represents a systemic vulnerability that neither the companies involved nor the governments responsible for national security can afford to ignore. The silicon triarchy of Broadcom, Nvidia, and the hyperscalers they serve is reshaping the architecture of AI power. The question of whether that reshaping serves the interests of broader humanity — or merely those of the few stakeholders who control the infrastructure — remains the defining governance challenge of the coming decade..
