Beyond the GPU: What Makes an AI Compute Platform Defensible?
Published · Dilikoglu Research
The strongest platform is the one that repeatedly lowers the cost of running reliable AI workloads. Shareholder value depends on how much of that benefit the supplier can retain.
Tugay Dilikoğlu | Information cutoff: 21 October 2026. Company disclosures and technical documentation are identified below; interpretations and the illustrative economics are Dilikoglu Research analysis.
Three investment takeaways
A faster accelerator creates potential value; software, networking and reliable deployment determine how much customers realize.
Switching costs vary by workload and customer capability. They can sustain an incumbent's returns without preventing a challenger from building a profitable business.
Platform durability should be tested through repeat deployments and cash returns after costs and reinvestment, rather than inferred from benchmark leadership or partnership scale.
Measure cost per useful workload
An accelerator can be cheaper to buy and more expensive to use. The relevant denominator is useful work delivered at the required quality, response time and reliability. Hardware price matters, but so do installation, software adaptation, energy, networking and operating support. A procurement saving disappears if it requires enough extra equipment or engineering to deliver the same service.
Consider a deliberately simplified comparison. Two systems have equal theoretical capacity. System A costs 100 units annually and delivers 70 units of useful output; System B costs 80 and delivers 50. Cost per useful unit is 1.43 versus 1.60. The lower-cost system is about 12% more expensive per useful unit. These are illustrative indexes, not measured NVIDIA or AMD results; costs include the same categories and output must meet identical service requirements.
Training and inference also require different comparisons. Training creates or updates a model; inference runs it for users. Time to reach an agreed training result is different from serving requests within a response-time limit. Even within inference, processing a long prompt and generating subsequent tokens can stress different resources. One performance result cannot establish superiority across the entire market.
What an integrated platform actually provides
NVIDIA's communications library, NCCL, coordinates work across GPUs within and between servers and supports several interconnect technologies. It illustrates why accelerator performance and system communication cannot be evaluated independently.1 The investment inference is that coordination can reduce the engineering customers must reproduce themselves. Documentation establishes functionality, however, not a quantified price premium or the cost of switching. NVIDIA's October 21 general-availability announcement for cuDSS, a library for sparse-system solvers, shows this software approach extending into engineering applications.10 Its vendor-reported application gains are workload-specific; they do not establish universal cost savings.
NVIDIA reported $89.0 billion of Data Center revenue and a 75.0% company-wide GAAP gross margin for its quarter ended 26 July 2026, designated fiscal Q2 2027.2 These results demonstrate commercial scale and current profitability. They do not isolate the contribution of software, networking, product mix or constrained supply. Treating the entire margin as a permanent software advantage would go beyond the disclosure.
Compatibility is a route, not a completed migration
AMD's HIP tools provide a path for converting CUDA source code. Its porting guide recommends incremental conversion and testing and distinguishes portable library interfaces from implementations optimized for AMD hardware.3 A path to functional compatibility therefore exists; equivalent production economics still need to be demonstrated for the intended workload. Rewriting every application from scratch and switching at no cost are both poor default assumptions.
There is useful customer evidence. Meta describes RCCLX, its communications work for AMD platforms, and a common Torchcomms interface intended to support different backends. It reports improvements in selected internal workload evaluations and explicitly limits some optimizations to single-node deployments.4 This shows that customers can help improve an alternative stack. It does not establish universal parity or make Meta's engineering costs representative of smaller buyers.
The resulting switching-cost question is specific: how much work must this customer perform, how often must it repeat that work, and how much economic benefit remains afterward? A large buyer can spread engineering expenditure over a substantial fleet. A smaller organization may rationally pay more for a familiar, supported system. Both decisions can be economically sound.
What benchmarks can tell investors
MLCommons' September 2026 MLPerf Inference v6.1 release added end-to-end retrieval-augmented generation and edge agentic tests, reflecting increasingly complex deployment patterns.5 The expansion itself is instructive: workload definitions evolve as applications change. Reproducible benchmark results are valuable evidence within those definitions, rather than a complete measure of every customer's operating cost.
Investors should compare the model, numerical precision, accuracy requirement, response-time constraint, hardware count and software configuration before accepting a performance ratio. A change in any of these may be economically relevant, but it changes the question being answered. A vendor's “up to” result should remain attributed to the vendor and its disclosed test conditions.
The decisive commercial test comes later: accepted equipment running customer workloads, followed by repeat deployment on attractive terms. Neither a benchmark entry nor an announced installation establishes utilization, collections or supplier returns. Public evidence rarely reveals all three, so a valuation should retain uncertainty where disclosure stops.
Durability framework — Dilikoglu Research analysis
Hardware and software coordination
Dependable workload economics
Adoption and repeat deployment
Supplier value capture
Delivery costs and reinvestment
Cash returns and next-generation capability
Conceptual framework. Technical improvements create an opportunity for customer value; adoption, bargaining and reinvestment determine the portion retained by shareholders.
The strongest case for a more open market
The challenge to incumbent returns is stronger than “another chip gets faster.” Customers can standardize interfaces, reuse deployment tools and direct selected workloads toward alternative accelerators. If those efforts reduce the cost of maintaining multiple suppliers across successive generations, an incumbent's historical ecosystem investment may become less effective protection for its future pricing.
AMD's October 12 Helion announcement extends its open-platform approach to agentic workloads, with CPU-only and integrated CPU/GPU configurations. Partner configurations are under development, with availability planned for November.6 This establishes a product roadmap, not accepted capacity or incremental revenue. Meta's engineering work provides a different form of evidence: customer effort to improve an alternative stack. Neither establishes fleet-wide cost parity.
The competitive field also extends beyond two GPU suppliers. Google announced both new TPU systems and NVIDIA-based instances at Cloud Next in April 2026.7 On October 21, NVIDIA said Google Cloud's A5X instances would become generally available in November, while naming an early-access customer.9 Early use and broader availability are distinct milestones. This portfolio approach supports coexistence of architectures; it does not establish equal economics or the displacement of GPUs.
An open market would not necessarily produce identical economics. Supplier differentiation could persist in particular workloads, support requirements or deployment environments. AMD can build a valuable business without displacing NVIDIA everywhere; NVIDIA can retain attractive returns without preserving every pricing advantage. What would weaken both is broad performance convergence combined with customers retaining most of the savings.
Evidence of durable returns
Layer
Customer benefit to test
Evidence that matters
Software
Less adaptation and debugging
Reproducible deployment and maintenance effort
System communication
More useful output from installed hardware
Performance at comparable cluster scale
Operations
Fewer interruptions and faster recovery
Sustained service performance, not peak throughput
Commercial support
Lower delivery and lifecycle risk
Accepted shipments and repeat orders
Supplier economics
Profitable share of customer value
Returns after support, incentives and reinvestment
Analytical framework, not a supplier ranking. Customer benefits and supplier returns must be established separately.
Value retained after the sale
More system content can increase revenue while adding integration work and warranty exposure. NVIDIA's October 21 supply-chain announcement includes a $1.1bn capacity reservation advance to KYEC, with high-volume production targeted for April 2027.11 Securing future supply can support growth, but cash is committed before that capacity produces returns. Software and supply assurance should not receive a separate valuation premium for benefits already embedded in the cash-flow forecast.
Commercial incentives require equal care. AMD's June 2026 filing describes warrants associated with its OpenAI and Meta agreements.8 These instruments can align incentives while transferring value to customers if conditions are met. Investors must account for customer incentives and potential dilution when translating revenue into shareholder value, without counting the same cost twice. Where contract terms are undisclosed, their effect on valuation remains uncertain.
Valuation depends on how long the supplier earns returns above its cost of capital. Repeat customer benefits can support that duration; stronger buyer bargaining can shorten it. Higher revenue, indefinitely elevated margins and a separate platform premium must not count the same advantage multiple times.
Evidence that would change the conclusion
Over the next two earnings releases: distinguish announced programs from delivered and accepted systems. Delivery progress strengthens commercial credibility but does not alone establish customer profitability.
With each comparable performance release: examine system configuration, accuracy and response-time requirements. Repeated gains on relevant workloads carry more weight than isolated headline ratios.
Over the next two to four quarters: look for customer accounts of migration effort, maintenance and repeat deployment. Lower recurring engineering requirements would strengthen the alternative-platform case.
Through the next product transition: track whether customers expand with the same supplier. Retention through a competitive procurement decision is stronger evidence than an initial evaluation.
Each reporting period: examine gross profit alongside development expense, inventory, receivables and disclosed incentives. Growth requiring disproportionate support may create less shareholder value than revenue suggests.
Over the next 12-24 months: test whether customer savings coexist with attractive supplier cash returns. Where public disclosure is insufficient, keep the conclusion conditional rather than substituting an assumed utilization rate or contract margin.
A defensible compute platform repeatedly makes demanding workloads easier or cheaper to operate and retains enough of that benefit to finance its next generation. Current evidence supports both the value of integration and a credible effort to make alternatives easier to adopt. The unresolved investment question is how the gains are divided as that competition matures. Investors should pay for demonstrated customer value and a credible duration of retained returns, while requiring the next deployment to test the thesis again.
Sources and evidence
NVIDIA, NCCL overview, version 2.32.3 documentation accessed 21 October 2026; publication date not stated. Technical functionality, not comparative customer economics.
NVIDIA, fiscal Q2 2027 results, 26 August 2026. Reported results for the quarter ended 26 July 2026; gross margin cited is consolidated GAAP, not a Data Center margin.
AMD, porting CUDA code to HIP, accessed 21 October 2026; publication date not stated. Technical migration guidance; the latest-documentation URL can change.
Meta Engineering, RCCLX on AMD platforms, 24 February 2026. Customer-authored engineering account and internal measurements, not an independent cross-vendor cost study.
MLCommons, MLPerf Inference v6.1, 16 September 2026. Benchmark scope and methodology; no supplier performance ranking is inferred here.
AMD, Introducing AMD Helion, 12 October 2026. Product announcement; partner configurations under development and November availability are management plans, not accepted shipments.
AMD, Q2 2026 Form 10-Q, filed 5 August 2026. Customer-warrant disclosures; quarter ended 27 June 2026.
NVIDIA, Google Cloud Vera Rubin infrastructure, 21 October 2026. Vendor announcement: general availability planned for November; early customer access does not establish broad deployment.
NVIDIA, cuDSS general availability, 21 October 2026. Software release and vendor-reported application evidence; no universal performance comparison inferred.
NVIDIA, supply-chain capacity expansion, 21 October 2026. Announced reservation advance and capacity plans; planned output is not existing productive capacity.
Disclosures
Tugay Dilikoğlu holds no NVIDIA or AMD securities, has no other material interest in either company, and has received no issuer or third-party compensation for the associated research. Meta and Google are cited as technical and customer examples; no investment recommendation on either is made.
Produced independently by Tugay Dilikoğlu, separately from his professional responsibilities. For informational purposes only; not personalized investment advice or an offer to transact. Opinions reflect information available through 21 October 2026 and may change. Company expectations and technical claims are not guarantees. This commentary does not initiate or change any company rating or price target. Readers should consult the underlying disclosures and consider their own circumstances.
Sign up through Beehiiv and confirm your email address to join. Beehiiv manages your subscription and unsubscribe options. Questions? research@dilikogluresearch.com.
The strongest platform is the one that repeatedly lowers the cost of running reliable AI workloads. Shareholder value depends on how much of that benefit the supplier can retain.
Tugay Dilikoğlu | Information cutoff: 21 October 2026. Company disclosures and technical documentation are identified below; interpretations and the illustrative economics are Dilikoglu Research analysis.
Three investment takeaways
Measure cost per useful workload
An accelerator can be cheaper to buy and more expensive to use. The relevant denominator is useful work delivered at the required quality, response time and reliability. Hardware price matters, but so do installation, software adaptation, energy, networking and operating support. A procurement saving disappears if it requires enough extra equipment or engineering to deliver the same service.
Consider a deliberately simplified comparison. Two systems have equal theoretical capacity. System A costs 100 units annually and delivers 70 units of useful output; System B costs 80 and delivers 50. Cost per useful unit is 1.43 versus 1.60. The lower-cost system is about 12% more expensive per useful unit. These are illustrative indexes, not measured NVIDIA or AMD results; costs include the same categories and output must meet identical service requirements.
Training and inference also require different comparisons. Training creates or updates a model; inference runs it for users. Time to reach an agreed training result is different from serving requests within a response-time limit. Even within inference, processing a long prompt and generating subsequent tokens can stress different resources. One performance result cannot establish superiority across the entire market.
What an integrated platform actually provides
NVIDIA's communications library, NCCL, coordinates work across GPUs within and between servers and supports several interconnect technologies. It illustrates why accelerator performance and system communication cannot be evaluated independently.1 The investment inference is that coordination can reduce the engineering customers must reproduce themselves. Documentation establishes functionality, however, not a quantified price premium or the cost of switching. NVIDIA's October 21 general-availability announcement for cuDSS, a library for sparse-system solvers, shows this software approach extending into engineering applications.10 Its vendor-reported application gains are workload-specific; they do not establish universal cost savings.
NVIDIA reported $89.0 billion of Data Center revenue and a 75.0% company-wide GAAP gross margin for its quarter ended 26 July 2026, designated fiscal Q2 2027.2 These results demonstrate commercial scale and current profitability. They do not isolate the contribution of software, networking, product mix or constrained supply. Treating the entire margin as a permanent software advantage would go beyond the disclosure.
Compatibility is a route, not a completed migration
AMD's HIP tools provide a path for converting CUDA source code. Its porting guide recommends incremental conversion and testing and distinguishes portable library interfaces from implementations optimized for AMD hardware.3 A path to functional compatibility therefore exists; equivalent production economics still need to be demonstrated for the intended workload. Rewriting every application from scratch and switching at no cost are both poor default assumptions.
There is useful customer evidence. Meta describes RCCLX, its communications work for AMD platforms, and a common Torchcomms interface intended to support different backends. It reports improvements in selected internal workload evaluations and explicitly limits some optimizations to single-node deployments.4 This shows that customers can help improve an alternative stack. It does not establish universal parity or make Meta's engineering costs representative of smaller buyers.
The resulting switching-cost question is specific: how much work must this customer perform, how often must it repeat that work, and how much economic benefit remains afterward? A large buyer can spread engineering expenditure over a substantial fleet. A smaller organization may rationally pay more for a familiar, supported system. Both decisions can be economically sound.
What benchmarks can tell investors
MLCommons' September 2026 MLPerf Inference v6.1 release added end-to-end retrieval-augmented generation and edge agentic tests, reflecting increasingly complex deployment patterns.5 The expansion itself is instructive: workload definitions evolve as applications change. Reproducible benchmark results are valuable evidence within those definitions, rather than a complete measure of every customer's operating cost.
Investors should compare the model, numerical precision, accuracy requirement, response-time constraint, hardware count and software configuration before accepting a performance ratio. A change in any of these may be economically relevant, but it changes the question being answered. A vendor's “up to” result should remain attributed to the vendor and its disclosed test conditions.
The decisive commercial test comes later: accepted equipment running customer workloads, followed by repeat deployment on attractive terms. Neither a benchmark entry nor an announced installation establishes utilization, collections or supplier returns. Public evidence rarely reveals all three, so a valuation should retain uncertainty where disclosure stops.
Conceptual framework. Technical improvements create an opportunity for customer value; adoption, bargaining and reinvestment determine the portion retained by shareholders.
The strongest case for a more open market
The challenge to incumbent returns is stronger than “another chip gets faster.” Customers can standardize interfaces, reuse deployment tools and direct selected workloads toward alternative accelerators. If those efforts reduce the cost of maintaining multiple suppliers across successive generations, an incumbent's historical ecosystem investment may become less effective protection for its future pricing.
AMD's October 12 Helion announcement extends its open-platform approach to agentic workloads, with CPU-only and integrated CPU/GPU configurations. Partner configurations are under development, with availability planned for November.6 This establishes a product roadmap, not accepted capacity or incremental revenue. Meta's engineering work provides a different form of evidence: customer effort to improve an alternative stack. Neither establishes fleet-wide cost parity.
The competitive field also extends beyond two GPU suppliers. Google announced both new TPU systems and NVIDIA-based instances at Cloud Next in April 2026.7 On October 21, NVIDIA said Google Cloud's A5X instances would become generally available in November, while naming an early-access customer.9 Early use and broader availability are distinct milestones. This portfolio approach supports coexistence of architectures; it does not establish equal economics or the displacement of GPUs.
An open market would not necessarily produce identical economics. Supplier differentiation could persist in particular workloads, support requirements or deployment environments. AMD can build a valuable business without displacing NVIDIA everywhere; NVIDIA can retain attractive returns without preserving every pricing advantage. What would weaken both is broad performance convergence combined with customers retaining most of the savings.
Analytical framework, not a supplier ranking. Customer benefits and supplier returns must be established separately.
Value retained after the sale
More system content can increase revenue while adding integration work and warranty exposure. NVIDIA's October 21 supply-chain announcement includes a $1.1bn capacity reservation advance to KYEC, with high-volume production targeted for April 2027.11 Securing future supply can support growth, but cash is committed before that capacity produces returns. Software and supply assurance should not receive a separate valuation premium for benefits already embedded in the cash-flow forecast.
Commercial incentives require equal care. AMD's June 2026 filing describes warrants associated with its OpenAI and Meta agreements.8 These instruments can align incentives while transferring value to customers if conditions are met. Investors must account for customer incentives and potential dilution when translating revenue into shareholder value, without counting the same cost twice. Where contract terms are undisclosed, their effect on valuation remains uncertain.
Valuation depends on how long the supplier earns returns above its cost of capital. Repeat customer benefits can support that duration; stronger buyer bargaining can shorten it. Higher revenue, indefinitely elevated margins and a separate platform premium must not count the same advantage multiple times.
Evidence that would change the conclusion
A defensible compute platform repeatedly makes demanding workloads easier or cheaper to operate and retains enough of that benefit to finance its next generation. Current evidence supports both the value of integration and a credible effort to make alternatives easier to adopt. The unresolved investment question is how the gains are divided as that competition matures. Investors should pay for demonstrated customer value and a credible duration of retained returns, while requiring the next deployment to test the thesis again.
Sources and evidence
Disclosures
Tugay Dilikoğlu holds no NVIDIA or AMD securities, has no other material interest in either company, and has received no issuer or third-party compensation for the associated research. Meta and Google are cited as technical and customer examples; no investment recommendation on either is made.
Produced independently by Tugay Dilikoğlu, separately from his professional responsibilities. For informational purposes only; not personalized investment advice or an offer to transact. Opinions reflect information available through 21 October 2026 and may change. Company expectations and technical claims are not guarantees. This commentary does not initiate or change any company rating or price target. Readers should consult the underlying disclosures and consider their own circumstances.