brain/
sourcestock-market

Autoresearch: Has TSMC or a major CPU vendor quantified the agentic-AI CPU wafer or revenue split since 16 Jul 2026?

Post-Q2 check: TSMC still will not size CPU vs GPU vs XPU; vendors sized their own CPU books, not TSMC wafers; CSP DRAM inventories not shown as normalized; later TSMC SEMICON commentary dropped the CPU-resurgence line.

Source

Autoresearch: Has TSMC or a major CPU vendor quantified the agentic-AI CPU wafer or revenue split since 16 Jul 2026?

Generated by /autoresearch on 2026-09-19. Synthesized across 3 rounds from 14 web pages (0 failed after one 10-Q retry), anchored by the Grokipedia TSMC entry. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market Research-only pass. No buy / sell / size. No invented TSMC CPU wafer percentage.

Summary

TSMC has not sized the CPU leg. On the 16 July 2026 Q2 call C.C. Wei again refused a CPU / GPU / XPU split and refused to update the AI CAGR. TSMC’s official IR calendar after that date lists only monthly-sales prints through January 2027; there is no later Wei investor conference as of 19 September 2026. The next scheduled-style earnings slot on the historical pattern is mid-October (Q3). A 31 August SEMICON Taiwan keynote by TSMC HPC BD director April Li kept “agentic AI” and token-volume language but dropped Wei’s “CPU resurgence, additive to accelerators” formulation.

Vendors sized their own CPU books, not TSMC wafers. After the TSMC call, AMD (4 August), Intel (23 July), NVIDIA (26 August call), and Arm (29 July shareholder letter) all kept an agentic-AI → server-CPU demand frame and published issuer growth or royalty figures. None of them, and no fetched TSMC primary, prints a CPU share of TSMC wafers or of TSMC revenue. NVIDIA’s 10-Q for the quarter ended 26 July 2026 does not disclose Grace/Vera as a line item; the $5 billion TTM Grace and “approximately $20 billion in total server CPUs” figures live on the call only.

US CSP DRAM inventories are not shown as normalized. The July TrendForce notes that sourced Step 2 still stand: server-CPU shortage slowed assembly and built DRAM at US CSPs in Q2, with CPU supply expected to improve from 2H26. Later TrendForce HTML (25 August, 7 September) does not reprint a CSP-inventory-cleared claim. Intel’s 23 July call is the strongest post-Q2 primary that the shortage persisted: server CPU demand “far outpace[s] available supply,” and management said they would “not catch up” even in Q4.

Findings

TSMC still will not size CPU vs GPU vs XPU

The load-bearing refusal is unchanged from the Q2 call. Asked by Citibank’s Laura Chen for an update on agentic-AI CPU growth versus GPUs and accelerators, Wei answered: “Laura, I don't think I can give you a very specific number, but let me share with you, all of them are in TSMC, and they're also using the same kind of leading-edge technologies. We working with our customer to allocate the wafer, the supply, to balance the CPU, GPU, XPUs ratio.” (Motley Fool TSMC Q2 transcript, 16 July 2026). Asked the same day whether the January mid-to-high-50% five-year AI-semi CAGR should be revised, he again declined a number: “If you asking about the AI's CAGR, let me give you not a number, but it's stronger and stronger. We don't give you the number today because it continue to increase, so we don't know how to answer this question. Stronger than what we said before.” (same transcript). Felix Pan asked whether the almost-$10 billion capex raise was “from CPU accelerator or memory components or back-end CoWoS”; the fetched HTML cuts off before Wei’s full answer, but the already-known call language attributes the raise to demand plus tool inflation, not a CPU percentage (same).

TSMC’s investor calendar after 16 July 2026 lists monthly sales only (10 August, 10 September, then 8 October / 10 November / 10 December / 8 January 2027). No later earnings conference or named Wei IR event is posted. Q2 itself sits in Past Events on 16 July; a Q3 2026 results date is not on the upcoming list as of this fetch (TSMC Financial Calendar). Official later Wei commentary that could keep or drop the CPU frame therefore does not exist in IR HTML.

Later TSMC commentary kept “agentic AI,” dropped the CPU-resurgence line

On 31 August 2026 (SEMICON Taiwan IC forum), April Li — TSMC’s AI and HPC business-development director, not Wei — said AI agents are moving into enterprise workflows and that inference token volume has increased “nearly 500-fold from its 2022 level,” with agentic systems consuming far more tokens than one-shot queries. Her quoted line is systems, not CPUs: “In this new era, market leaders will not be the ones who simply make better models, but those who build the most integrated systems.” She named logic, interconnect, memory, power/cooling, 3DFabric / CoWoS / COUPE, and HBM base dies. The fetched recap does not repeat Wei’s “resurgence in the role of CPUs” or “in addition to AI accelerators” (Taipei News / ANI via Focus Taiwan, 31 August 2026). That is a later TSMC voice on agentic AI that does not reaffirm the CPU-additive claim. It is also not a contradiction: silence is not a retraction. Wei himself has not been fetched saying he dropped it.

Grokipedia’s TSMC page is a historical primer (founding, nodes, Arizona, 2024–early-2026 financials). It does not cover the July 2026 agentic-CPU remarks and is not used as a 2026 sizing source (Grokipedia, TSMC).

Vendors kept the agentic-CPU frame and sized their books — still not a TSMC wafer split

AMD (4 August 2026 Q2 call; after TSMC). Lisa Su: “Looking ahead, agentic AI is creating a new growth vector for server CPUs, spanning high-frequency AI host nodes, high-density agentic servers and general-purpose cloud and enterprise workloads.” Fifth consecutive quarter of record server-CPU revenue; cloud and enterprise each “growing more than 70% year-over-year.” Data Center segment $6.7 billion, +107% YoY, 58% of company revenue — EPYC plus Instinct, not split. Company-level Q2 revenue $11.5 billion. Su guided “server revenue to grow more than 80% year-over-year in the second half of 2026 and more than 70% for the full year 2027” and a server-CPU TAM “more than 50% annually to approximately $220 billion by 2030.” Venice (Zen 6, TSMC 2 nm) “in production now”; demand “stronger than for any prior EPYC generation.” Asked for a CPU-versus-GPU dollar crossover into 2027, she refused a mix: “both businesses are going to grow a lot.” Asked which grows more on a dollar basis from Q2 to Q3, she again declined a split and said both grow, with more server supply in Q3/Q4 (Motley Fool AMD Q2 transcript, 11 August 2026 posting of the 4 August call). This is AMD server-CPU growth and an industry TAM, not a TSMC wafer percentage.

Intel (23 July 2026 Q2; seven days after TSMC). Lip-Bu Tan: “As AI expands from training to inference and increasingly to agentic and multi-agent systems, general purpose server CPU density continue to increase, and our core server CPU franchise is growing faster than ever.” “Demand accelerate across cloud and enterprise as customers increasingly recognize the critical role that CPUs in general and x86 CPUs, in particular, play in the AI infrastructure.” Q2 year-over-year server growth “the strongest on record.” Official IR: Q2 revenue $16.1 billion; Data Center and AI (DCAI) $6.3 billion — Xeon plus the rest of DCAI, not a CPU-only figure; Q3 guide $15.8–16.8 billion (Intel Q2 2026 results; Roic AI Intel Q2 transcript). On TAM, Tan said “the ratio of CPU to GPU going up” and “we now believe we're almost in parity at this point and could eventually even skew more to CPUs on a unit basis” — a unit comment, not a TSMC wafer or Intel revenue split, and not independently sourced (same transcript). Intel Foundry is the relevant capture path for Intel’s own Xeon (mostly Intel 3 / 18A), which cuts against a naive “almost all TSMC customers” read for x86 volume — already a known tension on the mechanism’s Step 3, not newly sized here.

NVIDIA (call ~26 August 2026 for Q2 FY2027 ended 26 July). Colette Kress / prepared remarks: “Rising adoption of agentic AI is driving an acceleration in demand for data center CPUs. Our Grace CPU, introduced in 2021, has been a great success, with revenue on a trailing 12-month basis exceeding $5 billion. Today, we are in full production of our next generation Vera CPU. As a standalone product, Vera expands our TAM even further.” “We continue to see demand for approximately $20 billion in total server CPUs. Based on our customer demand and improving supply outlook, our preliminary expectation is for CPU revenue to more than double in fiscal 2028.” Vera named “the first CPU built for AI agents” on the newsroom release; shipments to OCI, SpaceXAI, and AWS (starting that quarter) (NVIDIA newsroom, 26 August 2026; stockanalysis NVDA Q2 FY2027 transcript). Ambiguity to keep: “approximately $20 billion in total server CPUs” sits next to “our … CPU revenue to more than double.” It may be NVIDIA’s own demand book, an industry-wide figure, or something in between. Do not treat it as a TSMC wafer split. The Q2 FY2027 10-Q (period ended 26 July 2026, filed 26 August) reports Data Center $89.0 billion of $96.2 billion total and discusses Vera Rubin as the next architecture and inventory build; a grep of the fetched HTML finds no “Grace,” “Vera CPU,” “CPU revenue,” or “server CPU” dollar line (NVIDIA Form 10-Q). So the only NVIDIA CPU dollars in this pass are call commentary, not a 10-Q line.

Arm (shareholder letter for Q1 FYE27 ended 30 June 2026, filed 29 July — after TSMC Q2). Royalty revenue $715 million, +22% YoY, “with data center royalties more than doubling year over year.” Arm AGI CPU: secured capacity for the previously outlined $1 billion FY27–FY28 opportunity; “Customer demand has accelerated beyond our initial expectations and now exceeds $2 billion across fiscal 2027 and fiscal 2028.” Vera in full production (Arm-based); AWS–Meta “tens of millions of Graviton5 cores to power agentic AI workloads”; Google Axion as TPU host CPU; Microsoft Cobalt 200; Qualcomm Dragonfly C1000 (Arm Exhibit 99.2, SEC). These are royalty and Arm-silicon-opportunity figures, not a foundry wafer mix and not a TSMC disclosure.

What this does and does not do for Step 5. Vendors have now put dollars and growth rates on their CPU legs. That is new since 16 July. It still does not quantify TSMC’s CPU wafer or revenue share. Wei’s refusal stands. This pass does not invent a CPU % of TSMC wafers.

CSP DRAM inventories: not shown as normalized; Intel says the CPU shortage persisted

The Step 2 source is still the 9 July TrendForce HTML: “TrendForce notes server CPU shortages have slowed system assembly, leading to a gradual buildup of DRAM inventories at U.S. CSPs during the second quarter. CPU supply is expected to improve progressively from the second half of 2026 through 2027, allowing server production to ramp up.” CSPs “continue to increase procurement to prepare for future demand” into a 2027 server-DRAM shortage; 3Q26 server DRAM contracts +13–18% QoQ (TrendForce, 9 July 2026).

The 3 July TrendForce pricing note (six days before that pairing sentence, still useful for Q3 setup) says general-purpose x86 + RDIMM “remain the primary memory platform for Agentic AI workloads,” “Server shipments are expected to remain robust through 2027 as CPU availability continues to improve, supporting continued RDIMM consumption and inventory build-up during the second half of 2026,” and separately “CPU supply shortages have constrained system shipments” in enterprise SSD (TrendForce, 3 July 2026). Those two notes together say: shortage in Q2, expected H2 improvement, continued inventory build — not “inventories have normalized.”

Later HTML does not close the loop:

  • 25 August: memory (DRAM + NAND) estimated at 47% of major CSP capex in 2026 and 68% in 2027; server DRAM contract prices +~270% in 2026; no sentence that US CSP DRAM inventories normalized or that the CPU pairing shortage ended (TrendForce, 25 August 2026).
  • 7 September: 2Q26 DRAM industry revenue ~US$154.73 billion, +59.5% QoQ; “suppliers’ inventories remain at historic lows”; “The surge in agentic AI applications is further boosting RDIMM demand”; 3Q26 conventional DRAM still +13–18% QoQ. Silent on US CSP customer-side weeks and silent on whether the Q2 CPU-shortage pile-up cleared (TrendForce, 7 September 2026).

The strongest post-16 July primary that the server-CPU shortage was still binding is Intel on 23 July. Dave Zinsner: “server CPU demand continues to far outpace available supply.” Supply growth “more skewed towards the end of Q3 and into Q4, especially for servers.” “while things will improve, we will not catch up. We will be behind in the fourth quarter.” Lip-Bu: industry shortages across leading-edge wafers, memory, and substrates “will persist for the foreseeable future.” Outlook for industry server-CPU units: “strong double-digit unit growth … this year and next, with momentum extending into 2028” (Roic AI Intel Q2 transcript). AMD’s “improved supply” language is the vendor-side counterpart for EPYC, not a CSP-inventory print (AMD Fool transcript).

Read, carefully: Intel is a CPU supplier talking its shortage up; TrendForce’s later HTML never confirmed the Q2 CSP DRAM pile-up cleared. The honest status for Step 2 as of 19 September 2026 is: not demonstrated as normalized; Intel says the CPU shortage persisted into late Q3 / Q4; no new TrendForce HTML weeks figure.

Contradictions and open questions

  • TSMC CPU % of wafers remains unsized. Wei refused on 16 July. No later TSMC IR event. Vendors sized themselves. Do not back-solve a foundry mix from NVIDIA $5B / $20B, AMD $220B TAM, or Arm royalty doubling.
  • “$20 billion in total server CPUs” is ambiguous. Adjacent sentences mix NVIDIA CPU revenue (Grace TTM >$5B; FY28 CPU revenue more than double) with “total server CPUs.” Treat as call commentary, not a 10-Q line, and not a TSMC split.
  • Intel unit “parity” CPU vs GPU is a CEO remark on units, not dollars and not TSMC wafers. It sits next to Intel making most of its own server wafers — a live tension with Wei’s “almost all TSMC’s customers.”
  • Later TSMC voice dropped the CPU sentence without retracting it. April Li (31 August) is agentic + systems + packaging. Wei has not restated. Q3 call (typically mid-October) is the next place the frame can be kept or dropped on the record.
  • Step 2 is still one TrendForce pairing sentence plus Intel’s shortage. Later TrendForce HTML neither confirms nor denies CSP DRAM normalization. Paid TrendForce PDFs (3 September Server Market Bulletin; 9 September Server DRAM Industry Analysis) were not fetched (off the *.gov PDF whitelist). Their public blurbs still talk tight CPU/memory and “buyers … build buffer inventory”; those blurbs are not treated as fetched body text.
  • TSMC IR transcript PDF at investor.tsmc.com was not fetched (PDF, not *.gov). Fool HTML is the citation used for Wei’s wording.
  • No encyclopedia used as a 2026 sizing source. Grokipedia is vocabulary/history only.

Provenance

Rounds run: 3 (full)

Sub-questions by round:

Round 1 (broad survey):

  1. Has TSMC spoken after 16 July 2026 about agentic-AI CPU demand, and did later commentary keep or drop the additive-to-accelerators framing?
  2. Have Intel, AMD, Arm, or NVIDIA (Vera/Grace) quantified an AI-datacenter CPU wafer or revenue split since mid-July 2026?
  3. Have US CSP DRAM inventories normalized in Q3 2026, or has the server-CPU pairing shortage persisted?

Round 2 (drill-down):

  1. What exact NVIDIA Q2 FY2027 call language sizes Grace/Vera — is $20B NVIDIA demand or industry TAM? — targeting vendor dollars vs TSMC mix
  2. Did Intel’s 23 July call size Xeon vs accelerators or confirm a server-CPU shortage into Q3? — targeting Step 2 persistence
  3. Did later TSMC commentary (SEMICON) keep or drop Wei’s CPU-resurgence framing, and does later TrendForce HTML update CSP DRAM inventories? — targeting keep/drop + inventory

Round 3 (resolve remaining uncertainty):

  1. Does NVIDIA’s August 2026 10-Q disclose a CPU revenue line, or is Grace $5B / $20B call-only? — targeting whether Step 5 can move
  2. Does TrendForce’s 7 September HTML update US CSP DRAM inventories or the CPU pairing shortage? — targeting Step 2 close
  3. Confirm exact Wei Q2 refusal language from fetched HTML — targeting Step 5 quotes

Anchor source (Grokipedia, fetched before round 1):

  • TSMC — 12,017 chars extracted (capped; 196 citations; source: api) — historical primer (founding, nodes, Arizona, 2024 / early-2026 financials). Does not cover the 16 July 2026 agentic-CPU comments. Not used as a 2026 sizing source.

URLs fetched (14 successful, 0 failed after retry):

Round 1:

Round 2:

  • NVIDIA Q2 FY2027 transcript — earnings-transcript reprint (Reliable) — Grace TTM >$5B; ~$20B “total server CPUs”; FY28 CPU revenue more than double
  • Intel Q2 FY2026 transcript — earnings-transcript reprint (Reliable) — shortage persists; will not catch up in Q4; unit CPU/GPU “parity” remark
  • TrendForce, 3 July 2026 — industry press (Reliable) — x86+RDIMM as agentic platform; CPU availability “continues to improve”; H2 inventory build; SSD CPU shortage
  • Arm Exhibit 99.2 FYE27 Q1 — SEC primary (*.gov HTML) — DC royalties more than doubled; AGI CPU demand >$2B vs $1B secured capacity
  • AI growth entering 'industrialization' phase, TSMC — news recap of SEMICON Taiwan (31 Aug) — April Li; agentic/tokens/systems; no CPU-resurgence quote

Round 3:

Not fetched (noted, not cited as body text): TSMC IR Q2 transcript PDF (investor.tsmc.com, off *.gov PDF whitelist); TrendForce 3 September Server Market Bulletin PDF and 9 September Server DRAM Industry Analysis PDF; NVIDIA call PDF on carsongroup.com; SemiAnalysis / 404k Research (search-visible only).

Tools used: WebSearch, WebFetch, grokipedia.py fetch (TSMC). No --include-x. Generated: 2026-09-19 ~21:15 UTC

Referenced by
Mechanisms
Agentic AI → CPU resurgence in AI datacenters → additive silicon demand, ISA-agnostic → TSM captures it whoever wins; server-CPU scarcity is the corroborating tell