The AI-trade conversation, in four questions

What "Situational Awareness" predicted about stocks and funds

"Situational Awareness" refers to two related things by former OpenAI researcher Leopold Aschenbrenner: his June 2024 essay series "Situational Awareness: The Decade Ahead", and his AI-focused hedge fund Situational Awareness LP (launched July 2024). Note: it is a hedge fund, not a mutual fund, and it does not pick mutual funds — its only fund-type positions were in the VanEck Semiconductor ETF (SMH). Here are its predictions/bets, and how they turned out.

1. The essay's macro predictions (June 2024)

• AGI by ~2027, and true superintelligence by the end of the decade; AI outpacing many college graduates by 2025/26.
• Trillions of dollars pouring into GPUs, data centers and power ("the trillion-dollar cluster"); US electricity production growing by tens of percent.
• Heavy government/national-security involvement in AI by 2027/28; a US–China AI race.
• Implicitly bullish on the whole AI-compute supply chain — he mocked analysts who thought "2024 might be close to the peak" for Nvidia.

2. The hedge fund's stock bets (latest public filing: Q1 2026 13F, filed May 2026)

The fund disclosed $13.68 billion across 42 positions as of March 31, 2026 (SEC EDGAR, CIK 0002045724). Its implied predictions:

Bullish (long) — "second-order" AI infrastructure will keep re-rating:

• Bloom Energy (BE) ~$879M — power for data centers
• SanDisk (SNDK) ~$724M (plus $389M in calls) — storage/memory
• CoreWeave (CRWV) ~$556M (plus calls) — AI "neocloud" compute
• IREN ~$401M, Core Scientific (CORZ) ~$389M, Applied Digital (APLD) ~$320M, Riot Platforms, CleanSpark, Nebius, SharonAI, T1 Energy, HIVE Digital — data-center/power plays
• Calls on Micron (MU) $422M and TSMC (TSM) $355M
• Undisclosed internationally: big long positions in SK Hynix and Korean memory names; privately, a multibillion-dollar stake in Anthropic (bought at a $65B valuation in May 2026), whose IPO he predicted would be a major second-half-2026 catalyst.

Bearish (put options, ~$8.46B notional, 62% of the filing) — "first-order" AI-chip stocks are overpriced:

• SMH (VanEck Semiconductor ETF) $2.04B — a bet against the entire chip complex (the only "fund" prediction: semis ETF falls)
• Nvidia (NVDA) $1.57B, Oracle (ORCL) $1.07B, Broadcom (AVGO) $1.01B, AMD $969M, Micron $584M, TSMC $535M, ASML $494M, Intel $159M, plus small puts on Corning and Infosys
• Notably, Oracle was the only cloud name shorted (no Microsoft/Amazon/Google/Meta puts) — an implicit prediction that Oracle's AI-cloud story was over-promised. The fund also shorted legacy software seen as vulnerable to AI disruption (e.g., Adobe).

The overall thesis: AI demand is real and compounding, but the market was pricing it too narrowly through the big chip names, while power, data-center, memory and neocloud beneficiaries were still cheap — a "barbell": long infrastructure, short/hedged first-order chips and AI-disrupted software.

3. How the predictions turned out (as of early August 2026)

• For a while, spectacularly: the fund returned 439% net in the first half of 2026 (over 1,000% since inception), growing from a $225M seed (Stripe's Collison brothers, Nat Friedman, Daniel Gross; Jane Street an investor) to roughly $45 billion at its early-July 2026 peak.
• Then in July 2026 a historic "momentum crash" (Morgan Stanley's momentum index fell 17.4% in four days — the worst on record) hit both sides of its book: longs like Nebius, SanDisk, Micron, CoreWeave and SK Hynix fell 35–78%, while shorts like Adobe rallied. With reported leverage up to 400%, margin calls forced a deleveraging spiral.
• Per its investor letter, the portfolio sank ~67% in July. The fund was forced to sell its entire public stock book to Ken Griffin's Citadel at a discount, shrinking to roughly $10 billion (mostly the private Anthropic stake), and is unwinding — covered by CNBC, WSJ, Reuters, FT and NYT (July 30–31, 2026).

Bottom line

Situational Awareness predicted: AGI by ~2027, trillions into AI build-out, big gains for AI power/data-center/memory/neocloud stocks (Bloom Energy, SanDisk, CoreWeave, IREN, Core Scientific, Applied Digital, SK Hynix, Anthropic), and declines in richly-valued chip names and the SMH semiconductor ETF (Nvidia, AMD, Broadcom, Oracle, ASML, TSMC, Intel) and AI-vulnerable software (Adobe). Many of its macro calls on AI investment proved directionally right, but extreme leverage meant a single violent market reversal in July 2026 wiped out most of the fund before the thesis could play out — a modern echo of LTCM. Caveats: 13F filings are ~45 days stale, omit strikes/expiries/shorts/private holdings, so the puts may have been hedges rather than pure bearish predictions; none of this is investment advice.

Mutual funds to buy today, using "Situational Awareness"-style reasoning (as of early August 2026)

I researched current market conditions across 10+ sources (Reuters/WSJ/Business Insider coverage of the July 2026 momentum crash, SEC-derived fund-holdings data, Fidelity's official fund pages, Zacks/Yahoo utilities coverage, Korea/KOSPI crash analysis, and skeptical/bear commentary), and then reasoned from first principles rather than repeating anyone's list. Here is the logic, then the funds.

Step 1 — What the July 2026 crash actually proved (my reasoning, not the headlines)

• The thesis survived; the leverage didn't. Aschenbrenner's fund lost 67% in July with up to 400% leverage, yet is still +80% YTD. SK Hynix reported a 557% jump in quarterly operating profit the same week its stock was collapsing. Microsoft's earnings restored the rally (Nasdaq +2.8%), Amazon jumped 15% on results, and Meta is pouring nearly all free cash flow into AI capex. Conclusion: AI compute/power demand is intact; what failed was market structure (leverage, crowding, concentration).
• Forced-liquidation prices are informationless. As of July 31, SanDisk was still ~45% below its July 1 level, Bloom Energy ~30%, Lam Research ~30%, SK Hynix ~50% off its June peak — prices set by margin calls, not by new information about demand. Tellingly, Citadel outbid Millennium and Jane Street to buy SA's entire book: the most informed capital on Wall Street took the other side of the forced sale.
• Therefore the correct instrument is the one SA didn't use: an unlevered, diversified fund. The single biggest lesson of SA is that a right thesis dies if the wrapper can be margin-called. Mutual funds cannot be. This is the core of my recommendation: capture SA's (validated) second-order thesis in a wrapper that survives the next momentum crash.

Step 2 — Three SA-style edges that are actually available in mutual funds right now

1. The Anthropic NAV-lag. Dozens of ordinary mutual funds own pre-IPO Anthropic, and Q1-2026 filings show those stakes are still marked at the Series G-1 price — below the $965B Series H (May 2026) and far below Nasdaq Private Market's ~$689/share July estimate. With SA itself keeping only its Anthropic stake and an IPO expected in H2 2026, fund NAVs carry a small embedded, catalyst-driven mark-up that retail can buy at NAV with no premium.
2. Utilities are the only AI beneficiary with contractual cash flows. Hyperscalers (>$400B annual capex) are signing 10–20-year power purchase agreements; transformer and gas-turbine backlogs give power producers pricing power for years. Crucially, utilities were never part of the crowded momentum trade — they rose through July (XLU +8.6% YTD) and sit only ~10% off highs. This is the "picks-and-shovels of the picks-and-shovels": SA's Bloom/IREN/Core Scientific thesis, but in regulated, dividend-paying form.
3. The post-flush discount on second-order names. Memory, storage, and neocloud fundamentals (HBM demand, record profits) are intact while prices are 30–50% lower than six weeks ago. The right way to own this is diversified (a tech fund holding TSMC/Samsung/WDC among 100+ names), never single-name or single-country.

Step 3 — The funds

Tier 1 — Core: AI power & grid (35–40%)

• Fidelity Select Utilities (FSUTX) — $132.69, ~10% below its 2026 peak; ER 0.64%; top holdings NextEra 12.8%, Constellation Energy 6.6%, AEP 6.1% (nuclear/baseload-heavy, the direct data-center power play); +7.2% in Q1 2026, beat the S&P.
• Income-oriented alternatives: Franklin Utilities (FKUTX) or American Century Utilities (BULIX).

Tier 2 — Diversified growth with embedded pre-IPO Anthropic (25–30%)

• T. Rowe Price Capital Appreciation (PRWCX) — a balanced (stock/bond) fund with a $318M Anthropic stake (0.48% of AUM) and one of the best risk-adjusted records in the industry; the most conservative Anthropic access that exists. Caveat: it has historically been closed to new investors — if unavailable, use sibling PRCFX or the manager's ETF TCAF (no Anthropic in the ETF), or the next option.
• American Funds Growth Fund of America (AGTHX) — the single largest mutual-fund Anthropic position ($2.06B), inside a broadly diversified mega-cap growth portfolio.

Tier 3 — Tech funds with the highest Anthropic weight + post-flush memory exposure (15–20%)

• T. Rowe Price Global Technology (PRGTX) — 1.26% Anthropic plus TSMC 5.8% and Samsung Electronics 5.2%, i.e., diversified access to Korean/Taiwan memory at post-liquidation valuations; 1-yr +34.8% with lower volatility than semis funds.
• Fidelity Select Technology (FSPTX) — the highest Anthropic weight of any open-end mutual fund (1.42%), plus storage names (Western Digital 4.4%) that fit the second-order thesis; YTD +30.4%, 1-yr +43.1%, ER 0.61%, Morningstar 4★. Caveats: 21.6% NVIDIA (meaningful first-order exposure) and a brand-new manager as of 7/15/2026.

Tier 4 — Dry powder (10–15%)

• A money-market or short-bond fund (e.g., FXNAX/SPAXX). July proved that record momentum crashes now happen in days; unlevered cash is what lets you buy the next forced liquidation instead of being one.

What I would deliberately NOT buy — and why my reasoning differs from the crowd

• Fidelity Select Semiconductors (FSELX), despite (because of) its +76% 1-yr return: beta 2.13, NVDA ~21% + Broadcom ~13% + ASML — a concentrated bet on precisely the "first-order" layer SA was short, still the most crowded trade on earth even after the Philadelphia Semi Index fell 21% in July.
• Anything leveraged (the HK 2x SK Hynix product that ballooned to $7.8B was ground zero of the Korea crash) and single-country Korea funds — Samsung + SK Hynix are >50% of KOSPI; that concentration, plus Chinese DRAM entrant CXMT, is the real structural bear risk to memory margins.
• Destiny Tech100 (DXYZ) — 17.9% of NAV in Anthropic sounds appealing, but it habitually trades at extreme premiums to NAV; you'd pay several times fair value for the same exposure the mutual funds above give you at NAV.

Execution and honesty notes

• Dollar-cost average over 4–8 weeks. The rebound (VIX back to 16, S&P ~7,490) may not be the bottom; prime brokers are still tightening collateral on concentrated AI books, and "exhausted sellers are not repaired positioning."
• Fund holdings data is from Q1-2026 filings and June/July fact sheets and may have changed; Anthropic marks are periodic and opaque; a genuine AI-capex slowdown would hit every tier above except cash. Utilities carry interest-rate risk. None of this is investment advice — it is the portfolio that Situational Awareness's own logic implies once you remove the one thing that destroyed it: leverage.

Bottom line: ~35–40% FSUTX (or FKUTX), ~25–30% PRWCX/AGTHX, ~15–20% PRGTX (or FSPTX), ~10–15% cash — long the second-order AI build-out (power, grid, memory, storage) and pre-IPO Anthropic through unlevered, diversified wrappers bought at forced-liquidation discounts, while refusing the crowded first-order chip concentration and all leverage.

Should you buy SMH? My call: No — not as a new lump-sum buy today. Here's the reasoning.

I researched 10 current sources (VanEck's official fund page, stockanalysis.com, Reuters, WSJ/Barron's/MarketWatch headlines, Investopedia's July market coverage, and FXEmpire technicals) as of early August 2026. As with the earlier "Situational Awareness" analyses, I did my own reasoning rather than repeating anyone's rating. This is analysis, not personalized investment advice.

What SMH is right now (facts as of July 31 / Aug 1, 2026)

• Price ~$540.53, NAV $540.29; 52-week range $279.19–$671.83, with the all-time high on June 22, 2026 — so it sits ~20% below its peak after July's chip crash, but still +50% YTD and +83% over 1 year. The SOX index it mirrors more than doubled between late March and mid-June — a parabolic run.
• Valuation & structure: P/E ~44.7, beta 1.74, dividend yield 0.20%, expense ratio 0.35%, $68B AUM, only 26 holdings. Top-10 = 71.5% of assets: NVIDIA alone is 21.7%, TSMC 9.5%, Broadcom 6.7%, AMD 5.4%, ASML 5.1%, plus TXN, Micron, ADI, AMAT, QCOM.

Why I would not buy it today — five independent arguments

1. The "dip" is small relative to the run, and the price still embeds unbeatable expectations. A 20% pullback after a 100% three-month melt-up is a haircut, not a discount. The market is now selling record results: SK Hynix fell ~10% on a record $64B quarterly profit; Samsung fell after a 19-fold profit increase; an SK Hynix "miss" was defined as profit growing only +550% YoY. When the bar is that high at ~45x earnings on cyclical businesses near peak memory pricing, the payoff distribution is asymmetric to the downside.
2. The macro regime is hostile to 45x multiples. This is a rate-hike environment, not a cutting one: FedWatch puts September hike odds at 65%, three FOMC members dissented demanding action on inflation, the 10-year yield hit 4.75% (highest since Jan 2025), Brent is above $90 on the US–Iran war and the Hormuz blockade, and a third wave of tariffs is in force. Long-duration growth multiples compress when discount rates rise — SMH, at beta 1.74, is the pointy end of that.
3. The demand-side signal quietly flipped. The most telling event of earnings week: Microsoft was rewarded with a record +$450B day partly because capex came in lighter than expected. When the market starts paying hyperscalers for capex discipline, the growth rate of chip orders — SMH's entire earnings stream — is what gets sacrificed. Morgan Stanley also reports NVIDIA has already "de-specced" memory content of its racks "fairly materially." AI demand is real (AWS +37%, Apple forced to pay soaring memory prices), but SMH doesn't need demand to be real — it needs demand to keep accelerating to justify 45x.
4. Structural China risk hits SMH's exact holdings. In late July: CXMT completed an $8.6B Shanghai IPO and began mass-producing DRAM (threat to Micron, 4.7% of SMH); a Chinese state-backed firm reportedly began producing domestic DUV lithography machines (direct threat to ASML, 5.1%); Beijing is considering export controls on AI/chip tech; and China's Moonshot released an open-source model rivaling top US labs — the very report that cracked the sector on July 17. These aren't cyclical wobbles; they attack the scarcity premium that justifies the multiple.
5. Positioning and structure still look like a top, not a bottom. Retail piled in hardest in May (their biggest buying month of the year — at the highs); hedge funds sold chips four straight weeks; Investopedia's reporting notes "no one is short" — the crowding never cleared; Fitch says an AI correction is becoming a "major credit risk"; the options market says it's "not yet worth the risk to buy the dip"; and the rebound is a two-day spike off a July 29 low with negative breadth (decliners beat advancers even on Friday's up day, and SOX itself rose only 0.07% while Amazon ripped 15%). Technicians read this as a bear-market bounce toward SOX ~13,500 "before the next leg lower."

The honest bull case (why reasonable people are buying)

• Morgan Stanley's semiconductor analysts called the sell-off a "strong entry point" — shortages "show no signs of abating," memory prices are +25% this quarter, and NVDA/AVGO offer "the best risk-reward." JPMorgan says buy the summer dip. Intel beat on booming AI demand. AWS grew 37%. Washington signed $874M in new CHIPS letters of intent.
• SMH's long-run record is genuinely excellent: +34%/yr over 10 years. If AI capex keeps compounding, today's price will look cheap in 2030.

I don't dismiss this — but note the split: the stock-picking analysts are bullish while the strategists (Mike Wilson: semis correct through rest of 2026; BofA: semis seasonally weak in Q3) and the derivatives market are cautious. When the disagreement is about timing and multiple rather than demand, an index fund at 45x with 22% in one stock is the bluntest possible way to express the bull view.

Consistency check with the "Situational Awareness" framework

Recall from our earlier discussion: SA's single largest position was $2.04B of SMH puts — the thesis that first-order chip names were the overpriced, crowded layer of the AI trade, while power/memory/infrastructure were cheap. The July crash validated the direction of that view (SMH -20%, SOX -25%) even as the fund itself died of leverage. Nothing in the current data overturns the logic: SMH remains the most crowded, most expectation-laden, most China-exposed layer, now with a rate-hike Fed added on top. Buying SMH today is taking the exact side of the trade that the sharpest recent analysis was against — at a price only modestly below where that analysis was formed.

If you still want semiconductor exposure — a plan instead of a lump sum

• Size it small (≤5% of portfolio) and dollar-cost average over 8–12 weeks rather than buying at once — key events land soon: AMD earnings this week, NVIDIA earnings late August, the September Fed meeting, and the SOX 12,800–13,660 zone that technicians flag as the bounce ceiling.
• Accept that SMH has repeatedly drawn down 30–50% (it fell to $279 within the past 52 weeks); never touch the leveraged version (SOXL).
• Per the earlier analysis, more attractive expressions of the same AI build-out remain the unlevered second-order layer — utilities/power (e.g., FSUTX), diversified growth funds with pre-IPO Anthropic (AGTHX, PRGTX), and broad tech funds where TSMC/Samsung/memory are one slice among 100+ names — plus cash to buy the next forced liquidation.
• What would change my mind to "buy": SMH's P/E compressing toward the low-30s (via price or earnings catch-up), the Fed clearly off the hiking path, hyperscaler capex guidance re-accelerating (watch NVIDIA's data-center guide in late August), or a genuine positioning washout (real short interest, retail capitulation).

Bottom line

Don't buy SMH today as a fresh position. You'd be paying ~45x earnings for a 26-stock fund that is 22% NVIDIA, two days into a low-breadth bounce, in a rate-hiking, oil-shocked, tariff-ridden macro, where record profits are being sold, hyperscalers are being rewarded for spending less, and China is attacking the moat of its biggest holdings. The AI demand story is real — but at this price, SMH pays you for none of the risks. If you must own it, drip in small over 2–3 months; otherwise express the AI thesis through the cheaper, less-crowded second-order layer discussed previously. Sources: VanEck, stockanalysis.com, Reuters, WSJ, Barron's, MarketWatch, Investopedia, FXEmpire (data as of Aug 1–2, 2026; not investment advice).

Should you sell SMH and move into FSUTX, AGTHX, FSPTX? My call: Yes to rotating out of SMH — but not into this exact trio as constructed. Two of the three funds have flaws that partly defeat the purpose of the switch.

I pulled fresh data (as of July 31 – Aug 2, 2026) from 10 sources: stockanalysis.com pages for all four tickers, Fidelity's official FSPTX fund page, Capital Group's official AGTHX page, Yahoo Finance's FSUTX page, Morningstar's July 30 rotation article, and Google research on the Amundi/HSBC/VanEck August outlooks and data-center power demand (Gartner, Bloom Energy, Gabelli, Goldman). As before, this is my own reasoning, not a rehash of anyone's rating — and it is analysis, not personalized investment advice.

The four funds side by side (Jul 31, 2026)

              SMH        FSUTX       AGTHX        FSPTX
Price        $540.53    $132.69     $84.97       $52.24
Off 52wk hi  ~20%       ~10%        ~7%          ~11%
Beta         1.74       0.56-0.65   1.19-1.22    1.43
Portfolio PE ~44.7      (utility)   21.5         (tech, NVDA-heavy)
Semis weight ~100%      ~0%         ~13-16%      50.7%  <-- !!
NVIDIA       21.7%      0%          5.5%         21.6%  <-- !!
1-yr return  +83.0%     +15.4%*     +12.4%       +43.1%
July 2026    -17.6%     -4.3%       -3.5%        -6.1%
Yield        0.20%      ~5% ttm     0.25%        (low)
ER           0.35%      0.64%       0.59%+load   0.61%
Morningstar  n/a        5-star      3-star       4-star

*FSUTX total return per Yahoo/Morningstar (with reinvested distributions); price-only sources show lower. AGTHX figures are Class A.

Part 1 — Selling SMH: yes, the case is even stronger than when you asked "should I buy it"

• Everything from my previous "don't buy" analysis still holds: ~45x earnings on cyclical businesses, 21.7% in one stock, a rate-hiking Fed, oil at ~$85–91 on the Hormuz crisis, China's CXMT/DUV attacks on the moat, record profits being sold (SK Hynix −10% on a record $64B quarter), and hyperscalers being rewarded for spending less. Holding is a milder form of the same bet as buying.
• The institutions have now confirmed the rotation. In the past week alone: Amundi's August outlook says "the rotation theme is confirmed... out of chips"; HSBC's August monthly documents the rotation from semis/hardware into hyperscalers and software; hedge funds sold chips four straight weeks; WSJ ran "Investors Rotate Out of Chip Stocks." The Morningstar Global Semiconductor Index gained 60% in H1 and is still 17% below its June peak.
• But don't sell 100% in one order. SMH is two days into a rebound (VIX back to 16, Nasdaq +1%), JPMorgan and Morgan Stanley's chip analysts are still bullish, and NVIDIA reports in late August. A staged exit (e.g., 25–33% per tranche over 4–8 weeks, or sell down to a small 10–15% residual) captures the de-risking without betting everything on one timing decision.

Part 2 — The destination trio: one great fit, two fixable problems

FSUTX (Fidelity Select Utilities) — the strongest leg. Keep it, make it the anchor.

• The structural story got stronger this month: Gartner now sees data-center electricity demand up 26% in 2026 (565 TWh); S&P GMI sees US data-center power demand going 60→80+→180 GW (2025→2026→2030); 23 states have approved large-load tariffs with 14-year contracts and per-MW collateral — i.e., contractual revenue for exactly FSUTX's holdings (NextEra 12.8%, Constellation, AEP, Vistra, NRG).
• Morningstar 5★, beta ~0.6, ~10% below its high, 3-yr return +17.3%/yr vs category +9.2%. It fell only 4.3% in July's crash vs SMH's 17.6% — exactly the diversification you're buying.
• Caveats: it is rate-sensitive (a hiking Fed is a headwind — that's partly why it's off its high), non-diversified, and 76% turnover means real capital-gains distributions (Dec/Apr).

AGTHX (Growth Fund of America) — right idea, possibly the wrong share class.

• The concept is sound: 333 holdings, beta 1.19, portfolio P/E 21.5 (vs SMH ~45), downside capture 85, plus the largest mutual-fund pre-IPO Anthropic stake ($2.06B) ahead of an expected H2-2026 IPO — the "second-order AI" exposure from our earlier discussion.
• But Class A (AGTHX) carries a 5.75% maximum front-end sales load. Capital Group's own page confirms it. Some brokerages waive it; many don't. Paying 5.75% upfront would erase years of expense advantage. Before buying, confirm your platform sells it load-waived; if not, use F-3 (GAFFX) through an advisor platform, or a no-load large-growth fund with similar character. Also note it's only a 3★ fund with active share of 48 — it hugs the index; you're mostly buying diversification + Anthropic, not manager brilliance.
• Overlap check: it still holds ~13–16% in chips (NVDA 5.5%, AVGO 4.4%, MU 3.3%, TSMC 2.9%) — acceptable dilution, but know that it isn't chip-free.

FSPTX (Fidelity Select Technology) — the flawed leg. This is where I'd push back hardest.

• FSPTX is 50.7% semiconductors and holds MORE NVIDIA (21.6%) than SMH itself (21.7% — effectively identical). Its top-10 is NVDA, Apple, Marvell, Cisco, NXP, ON Semi, Western Digital, GlobalFoundries, Broadcom, Astera Labs. Selling SMH to buy FSPTX is, for half of every dollar, selling your chip fund to buy a different chip fund — at 0.61% ER instead of 0.35%.
• It also just got a brand-new primary manager (Priyanshu Bakshi, since 07/15/2026) — two weeks of tenure through the most violent semi market in years.
• If you want tech-that-isn't-SMH, the current rotation evidence points elsewhere: Morningstar, ING, and Deutsche Bank all document money moving into software (trading at a 6% premium to the S&P vs a 36% historical average, with +24% expected 2026 earnings growth — DB explicitly calls it "a good hedge against volatility in semis"). A diversified global tech fund like PRGTX (which Fidelity itself lists as a comparable, with TSMC/Samsung as slices among many rather than a 22% NVDA block, plus a 1.26% Anthropic stake) or a software-tilted fund expresses "tech without the SMH concentration" far better. If you keep FSPTX at all, size it small and treat it as continuing chip exposure, not diversification.

The math that matters: how much de-risking do you actually get?

• Equal thirds into FSUTX/AGTHX/FSPTX takes you from 100% semis → ~22% semis and from 21.7% NVIDIA → ~9% NVIDIA, with blended beta falling from 1.74 to roughly 1.1. That is a real improvement — the plan is directionally right even as proposed.
• But swapping the FSPTX third for PRGTX-or-software (or simply more FSUTX + cash) takes semis to ~7–10% and NVIDIA to ~3%, which is what "getting out of SMH" should actually mean.

Taxes and mechanics — check these before pressing sell

• Account type decides everything. In an IRA/401(k), rotate freely. In a taxable account, SMH is +83% over one year and +50% YTD — selling realizes a large gain. If any lots are short-term, waiting for long-term treatment may be worth more than the timing risk; consider spreading sales across the 2026/2027 tax years and harvesting any losers elsewhere against the gain.
• Don't jump in front of distributions. Actively-managed funds pass out taxable capital gains: AGTHX distributed $8.60/share (~10% of NAV) last December; FSPTX paid $1.41 in April and $2.92 in November; FSUTX pays Dec/Apr. In a taxable account, buying shortly before a distribution date means paying tax on gains you didn't enjoy — check each fund's estimated distribution calendar (Oct–Dec) before buying large amounts in Q4.
• Ongoing tax drag: Fidelity's own data shows FSPTX's 1-yr return drops from 59.8% to 56.0% after taxes on distributions. ETFs like SMH are structurally more tax-efficient; you are trading tax efficiency for diversification — fine, but do it knowingly.
• Structural upgrade to consider: in a taxable account, the ETF versions of these exposures (e.g., a utilities ETF instead of FSUTX, a diversified growth ETF instead of AGTHX-with-load) deliver nearly identical portfolios with no loads and minimal distributions.

A concrete plan (if your risk tolerance matches the earlier discussions)

• Sell SMH in 3–4 tranches over 4–8 weeks (key events: AMD and NVIDIA earnings in August, the September Fed meeting). Optionally keep a 10–15% residual so a renewed chip melt-up doesn't leave you fully out.
• Redeploy roughly: ~40% FSUTX (anchor; contractual AI-power cash flows, 5★, low beta), ~25–30% AGTHX only if load-waived (else GAFFX/no-load equivalent) for diversified growth + Anthropic, ~15% PRGTX or a software-tilted tech fund in place of most of FSPTX (0–10% FSPTX only if you consciously want to stay ~50% in semis within that sleeve), and ~15% money-market/short bonds — July proved the value of unlevered dry powder at 4%+ yields.
• What would change my mind: if SMH's P/E compresses to the low-30s, the Fed pivots off the hiking path, or hyperscaler capex guidance re-accelerates at NVIDIA's late-August report, the case for keeping more SMH strengthens; conversely a September hike or another China breakthrough argues for accelerating the exit.

Bottom line

Yes — rotating out of SMH is the right move and consistent with everything in our prior analyses (SMH remains the most crowded, most expectation-laden, most China-exposed layer of the AI trade at ~45x earnings in a hostile macro). FSUTX is an excellent destination and should be the largest leg. But fix two things first: (1) don't buy AGTHX Class A if it costs you the 5.75% load — use a load-waived/no-load route; and (2) recognize that FSPTX is half semiconductors with a 22% NVIDIA stake and a two-week-old manager — it recreates much of what you're selling, so shrink or replace it (PRGTX or a software-leaning fund fits the current rotation evidence better). Execute in stages, mind the capital-gains hit and Q4 distribution dates in taxable accounts, and keep some cash. Sources: stockanalysis.com, Fidelity, Capital Group, Yahoo Finance/Morningstar, Amundi, HSBC, VanEck, Gartner, WSJ/Reuters coverage (data as of Jul 31–Aug 2, 2026). Not investment advice.