AMD Advanced Micro Devices

Research status: Re-checked through 25 Sep 2026. Valuation reference: $629.26 close on 24 Sep 2026.

The key September conclusion is not a new fundamental breakthrough; it is a valuation reset. The August dossier already captured the core Helios/OpenAI/Meta/Anthropic evidence, customer warrants and capital commitments. Since then, the stock has rerated from $473.25 to $629.26 (~33%) without a comparably large new public proof point on production acceptance, ROCm depth, customer economics or fully diluted per-share returns.

AMD remains strategically impressive: Data Center revenue was $6.7B in Q2 (+107% YoY), OpenAI and Meta each target up to 6GW, Anthropic up to 2GW, and Helios is intended to begin coming online with OpenAI in Q4 2026. But the investment hurdle is now materially higher. The prior 24-month PowerFund scenario range—bear $260–330, base $500–600, bull $750–900—still looks directionally usable pending Q3. At $629, the stock is already above the prior base range midpoint and above the old probability-weighted range of roughly $500–590.

PowerFund therefore keeps AMD at investigate / wait for stronger deployment economics or a material dislocation. Do not raise fair value merely because the market capitalization crossed $1T. The next legitimate upward re-underwrite requires evidence that Helios/customer economics, gross profit and ROCm adoption are improving enough to offset warrants, capital commitments and dilution.

Investment case

AMD has become the most credible second merchant full-stack AI compute platform, combining EPYC CPUs, Instinct GPUs, Helios rack-scale systems, Pensando networking and ROCm software.

> The PowerFund thesis remains that AMD can create substantial per-share value if Helios establishes repeatable production economics across multiple large customers while EPYC continues to gain share. At the current valuation, however, the burden of proof is much higher: customer commitments must convert into durable gross profit and per-share FCF that more than compensate for warrants, investments, lease support and other strategic concessions.

September valuation reset

• Price: $629.26 on 24 Sep 2026 versus $473.25 at the August underwriting reference.
• No comparably large new public evidence has yet emerged on production acceptance, independent Helios TCO, customer-level gross profit or warrant-adjusted EPS.
• Existing 24-month scenario framework remains approximately:
• Bear: $260–330
• Base: $500–600
• Bull: $750–900
• At the current price, expected return is now heavily dependent on the bull case or a material upward revision to normalized earnings power after Q3.

What would justify a higher value

1. Q3/Q4 Data Center revenue and margin materially exceed the current framework without gross-margin sacrifice.
2. OpenAI/Meta/Anthropic deployments move from announced commitments to accepted production at attractive economics.
3. ROCm shows independent production adoption beyond customers receiving strategic incentives.
4. Warrant/commitment economics prove clearly accretive on a fully diluted per-share basis.
5. Helios demonstrates competitive production TCO versus NVIDIA alternatives, not just vendor benchmarks.

• Q3 2026 delivery above the $13.0B ± $0.3B revenue framework without gross-margin erosion.
• OpenAI Helios deployments beginning in Q4 2026 and accelerating during 2027.
• Meta validation converting to large-scale production deployment.
• Anthropic's first MI450/Helios gigawatt beginning on schedule in H1 2027.
• Named deployments at Microsoft Azure, Oracle and other CSPs broadening customer concentration.
• MI450 / Helios benchmark and real production TCO evidence versus NVIDIA Rubin-class systems.
• ROCm production adoption expanding beyond customers with strategic/economic incentives.
• EPYC Venice / 6th Gen share gains continuing independently of GPU wins.
• Data Center operating margin remaining ≥30% as GPU mix rises.
• Commitments and warrant dilution proving accretive to per-share earnings/FCF.

• Valuation/crowding is now the immediate risk: the stock has rerated ~33% from the August underwriting mark.
• Warrant dilution: OpenAI and Meta each hold up to 160M-share warrants at $0.01 if milestones are met.
• Strategic commitments: ~$30.3B unconditional commitments, $4.1B partner lease guarantees, up to $5B Anthropic investment and ~$9.5B future leases remain material.
• NVIDIA moat: CUDA/networking/libraries/cluster operations remain formidable.
• Customer concentration: early Helios economics depend heavily on a small group of strategic customers.
• ROCm execution: production reliability and software-porting burden remain key uncertainties.
• Custom ASIC competition: hyperscalers can internalize more inference/training workloads.
• AI-capex factor concentration: AMD would add strongly to PowerFund's dominant factor.

Warning — freeze additions and re-underwrite

• Q3 revenue below $12.7B or non-GAAP gross margin below 55%.
• Data Center growth falls below approximately 35% YoY before anchor Helios deployments are established.
• Data Center operating margin falls below 28% for two quarters.
• OpenAI/Meta first meaningful Helios deployments slip beyond Q1 2027 or Anthropic's first gigawatt slips beyond Q3 2027.
• Customer-related warrants/commitments grow materially faster than the revenue/earnings they are supporting.
• ROCm production adoption remains concentrated in counterparties receiving significant strategic economics.

Reduce

• Data Center revenue declines sequentially for two quarters while broader AI-capex demand remains healthy.
• FY2027 Data Center growth trajectory falls below ~30% without a clear temporary supply constraint.
• Company non-GAAP gross margin falls below 54% by mid-2027 despite richer Data Center mix.
• Fully diluted share count rises materially without proportional increase in base-case per-share value.
• Helios total-cost-of-ownership evidence is consistently >15% worse than comparable NVIDIA systems for target production workloads.
• Commitments/guarantees begin producing material losses or cash calls.

Invalidate

> Invalidate if AMD fails to establish Helios/ROCm as a repeatable, economically attractive second AI platform—specifically, if anchor deployments slip or remain subsidy-driven, Data Center growth falls toward ordinary semiconductor rates, and warrants/commitments dilute per-share economics faster than earnings compound.

Additional hard invalidators include sustained loss of EPYC server share, a major MI-series reliability/qualification failure, or multi-generation roadmap slippage that leaves AMD structurally behind NVIDIA/custom accelerators.

AMD is the most credible second merchant full-stack compute platform because it combines server CPUs, GPUs, networking, rack-scale systems and an increasingly mature open software stack.

• NVIDIA: dominant in accelerator share, CUDA, networking, libraries and cluster operations. AMD does not need to displace NVIDIA; it needs enough economically attractive share to make the huge AI market multi-sourced.
• Hyperscaler ASICs: Google, AWS, Microsoft and others can remove merchant GPU demand from stable/high-volume workloads, especially inference.
• Intel / alternative merchant compute: currently weaker in leading AI accelerators but remains relevant in CPUs/foundry ecosystems.

AMD's strongest structural advantages are EPYC's proven server franchise, chiplet expertise, high-memory configurations, TSMC access and customer appetite for an open second source. Its weakest layer remains software/ecosystem depth versus CUDA.

The OpenAI/Meta warrants are strategically ambiguous: they increase customer commitment but also demonstrate that AMD shared meaningful economics to secure anchor deployments. PowerFund therefore treats per-share return after incentives as the core moat test.

1. After Q3 2026, rebuild the customer deployment table for OpenAI, Meta, Anthropic, Microsoft, Oracle and other Helios adopters: contractual status, first delivery, production acceptance, revenue timing and incentives.
2. Model OpenAI/Meta warrant dilution tranche by tranche against purchase milestones and incremental gross profit / EPS.
3. Reconcile the $30.3B unconditional commitments, $4.1B lease guarantees, up-to-$5B investments and $9.5B future leases by counterparty and strategic purpose.
4. Track Data Center gross profit and operating margin, separating EPYC from Instinct/Helios where disclosure permits.
5. Gather independent Helios production TCO/utilization evidence versus NVIDIA rack-scale alternatives.
6. Track ROCm production workloads and software-porting burden outside strategically incentivized anchor customers.
7. Re-underwrite 24/60-month value only when normalized earnings power changes materially; do not anchor to market cap milestones.

Next full review: after Q3 2026 results, or earlier on a material OpenAI/Meta/Anthropic production acceptance or economic-disclosure event.